Examenos

Decision Index public suite (v0.3) as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

DecisionBench (Hanno Labs) v1 as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Decision Bench (decisionbench.ai) v4 as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Fast Decisions as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

JevBench v1.4.1 as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

ImajevBench v2.0-lite as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

WorkflowEvals as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Typed Decisions as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Liquid decision board (over a hundred public and private benchmarks) as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Decisions.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

AI2D as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

BLINK as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

CountQA as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

CV-Bench as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

ERQA as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

HallusionBench as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

MMBench as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

MME as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

MMMU-Pro as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

MMRB2 as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

MMStar as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

MMVP as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

NaturalBench as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

POPE as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

V* as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

VisualPuzzles as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

VisualWebBench as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

VL-RewardBench as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Average (vendor mean of shared public benchmarks) as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the LFM3 d1 card benchmark table; unpublished cells are omitted, not zero.

Liquid AI vendor measurements from the LFM3 d1 model cards; the d1-2B and d1-Omni-2B cards publish the identical four-model table. All four models are scored on the same items with the same scoring, so a few d1-3B and d1-omni-600M values differ slightly from their own October cards. Average is the vendor's unweighted mean of the public benchmarks all four models have (excludes Fast Decisions and the board). Image benchmarks are read as decisions over each benchmark's options, on at most 1,000 rows each; Liquid calls POPE, MME and VisualWebBench saturated. The card does not name each benchmark's metric; values are kept as printed. Exact item populations, prompts and served precision are unpublished. Card group: Vision.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Spoken questions, on tasks absent from training · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Spoken questions, on tasks absent from training · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Spoken short states · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Spoken short states · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Assistant commands, real speakers (SLURP intents) · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Assistant commands, real speakers (SLURP intents) · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Assistant commands in 12 languages (Speech-MASSIVE intents) · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Assistant commands in 12 languages (Speech-MASSIVE intents) · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Spoken language, 43 languages (FLEURS) · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Spoken language, 43 languages (FLEURS) · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Sound events (ESC-50) · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Held-out audio.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · DecisionBench (Hanno Labs) v1, routing and triage · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · DecisionBench (Hanno Labs) v1, routing and triage · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · JevBench v1.4.1, the rows a voice can say · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · JevBench v1.4.1, the rows a voice can say · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Abstention: none of the options applies (our set) · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Abstention: none of the options applies (our set) · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Typed Decisions, customer service · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Typed Decisions, customer service · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Belebele, passages in 8 languages · written as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Audio · Belebele, passages in 8 languages · heard as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the d1-Omni-2B card audio table.

Liquid AI vendor measurement on the d1-Omni-2B card. Each row is scored with the question written and heard, on clips and voices absent from training; effective accuracy where 0 is chance and 100 is every answer right. In the read-aloud group, synthetic voices read the question and short state of 1,284 test rows in each row's language. Populations other than stated are unpublished. Card group: Decision suites read aloud.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · NVIDIA H100 PCIe · One question as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · NVIDIA H100 PCIe · 3 questions as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · NVIDIA H100 PCIe · 5K-token state as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · NVIDIA H100 PCIe · 384 px picture as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · NVIDIA H100 PCIe · 10 s audio clip as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · NVIDIA H100 PCIe · 64 states in one call (states/s) as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · AMD MI325X · One question as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · AMD MI325X · 3 questions as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · AMD MI325X · 5K-token state as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · AMD MI325X · 384 px picture as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · AMD MI325X · 10 s audio clip as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · AMD MI325X · 64 states in one call (states/s) as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · Apple M5 Pro (mps) · One question as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · Apple M5 Pro (mps) · 3 questions as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · Apple M5 Pro (mps) · 5K-token state as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · Apple M5 Pro (mps) · 384 px picture as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · Apple M5 Pro (mps) · 10 s audio clip as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified

Speed · Apple M5 Pro (mps) · 64 states in one call (states/s) as reported by Liquid AI on the LFM3 d1 model cards

Released open d1-Omni-2B checkpoint, card revision c8d48c08dada0191c6b362820f055cfb49c99426.

From the model card speed table; vendor timing, no score ranking.

Liquid AI vendor timing: warm calls after model.compile() in bf16, median of 20 calls, best of three runs on the H100 and the M5 Pro. Hardware and setup are vendor-specific; not a hosted-API latency or tariff.

Checked / reported: 2026-10-28 · Samples: Not published

Snapshot: 2026-10-27 · Revision: d1-2B cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41; d1-Omni-2B c8d48c08dada0191c6b362820f055cfb49c99426 · Published observations; result reuse terms not specified