Examenos

27.7% · raw 27.69 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.6, top_p 0.95, max_turns 50 per official HF card, output bounded by context. Version/public split not stated; do not infer v1.0.6 or AA private reward configuration. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

65% · raw 65.01 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.001, maximum output 4,096 tokens; overall as labeled, not multi-turn-only. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

80.7% · raw 80.71 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0, maximum output 1,024 tokens; GPT-4o-2024-08-06 user simulator at temperature 0 and maximum 1,024 tokens. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

48.1% · raw 48.09 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM default temperature 1.0 (agent sets no temperature), maximum output 8,192 tokens; exact suite revision not stated. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

83.2% · raw 83.18 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Official card groups unlisted evaluations under temperature 0.6 and maximum output 32,768 tokens; specific PinchBench harness settings/revision are not detailed. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

8% · raw 8.04 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Official GPT-5.2 user model at low reasoning effort, default BM25 retrieval; agent temperature 0, maximum output 16,384 tokens per turn. The user simulator effort is not the evaluated model effort. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

26.8% · raw 26.75 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.6, top_p 0.95, max_turns 40 browser actions; output bounded by context. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

74.4% · raw 74.42 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.6, maximum output 32,768 tokens. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

-24.9 · raw -24.9 index

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Raw signed vendor-reported AA-Omniscience score, not accuracy or non-hallucination rate and not an independently measured AA result. Split, normalization and exact scoring formula are not specified; do not blend with explicitly public-split scores. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

7.4% · raw 7.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Source labels this HLE and groups it with temperature 0.6 / maximum output 32,768 tokens. Exact subset and tool protocol are not explicitly documented, so keep separate from full/text-only/with-tools HLE configurations. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

96.8% · raw 96.75 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.6, maximum output 32,768 tokens. Strict/loose aggregation is not specified. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

74.7% · raw 74.69 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.6, maximum output 32,768 tokens; source labels IFBench without a strict/loose qualifier. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

84.4% · raw 84.38 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.6, maximum output 32,768 tokens; exact language/turn aggregation unpublished. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

96% · raw 96.04 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.6, maximum output 32,768 tokens; sample aggregation unpublished, so do not assert pass@1 or average count. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

51% · raw 51 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. OpenCode agent harness explicitly named; temperature 0.6 and maximum output 32,768 tokens per card catch-all settings. Keep separate from other SWE agent harness configurations. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

32.7% · raw 32.71 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. vLLM; temperature 0.6, maximum output 32,768 tokens; vendor measurement, not independent AA intelligence-index data. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

4318 tok/s · raw 4318 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; Apple M5 Max, llama.cpp, Q4_K_M quantization, 4K-token input. prefill transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

184 tok/s · raw 184 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; Apple M5 Max, llama.cpp, Q4_K_M quantization, 4K-token input. decode transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

15 GB · raw 15 GB

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; Apple M5 Max, llama.cpp, Q4_K_M quantization, 4K-token input. memory transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

2451 tok/s · raw 2451 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; Mac Studio (chip unspecified), llama.cpp, Q4_K_M quantization, 4K-token input. prefill transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

146 tok/s · raw 146 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; Mac Studio (chip unspecified), llama.cpp, Q4_K_M quantization, 4K-token input. decode transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

16 GB · raw 16 GB

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; Mac Studio (chip unspecified), llama.cpp, Q4_K_M quantization, 4K-token input. memory transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

1641 tok/s · raw 1641 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; AMD Ryzen AI Max+ 395, llama.cpp, Q4_K_M quantization, 4K-token input. prefill transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

114 tok/s · raw 114 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; AMD Ryzen AI Max+ 395, llama.cpp, Q4_K_M quantization, 4K-token input. decode transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

16 GB · raw 16 GB

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI vendor evaluation of the released reasoning-only checkpoint. Pipette; AMD Ryzen AI Max+ 395, llama.cpp, Q4_K_M quantization, 4K-token input. memory transcribed from the labeled official figure. Device reasoning budget/concurrency not detailed. No cloud API speed inference. Evaluation date unpublished; source checked 13 October 2026. No selectable effort level is specified.

21070 tok/s · raw 21070 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 1024. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

17940 tok/s · raw 17940 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 512. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

13890 tok/s · raw 13890 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 256. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

9659 tok/s · raw 9659 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 128. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

6131 tok/s · raw 6131 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 64. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

3794 tok/s · raw 3794 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 32. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

2347 tok/s · raw 2347 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 16. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

1484 tok/s · raw 1484 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 8. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

887.7 tok/s · raw 887.7 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 4. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

538.1 tok/s · raw 538.1 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 2. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.

341.5 tok/s · raw 341.5 tok/s

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-13

Harness: Liquid AI, one NVIDIA B300, BF16, SGLang 0.5.22; 1,024 input tokens / up to 256 output tokens, sustained concurrency 1. Total output tokens divided by wall time, continuously replacing completed requests. torch.compile where SGLang allows; default settings with per-model serving adjustments. Exact published value from official interactive Figure 5 tooltip/accessible table, not digitized pixels. Peak 1,024-concurrency result is headline; ten alternatives retained. Reasoning budget unspecified; checked 13 October 2026, evaluation date unpublished.