Clover-35B-A3B-Preview Benchmarks, Specifications & Availability

Clover-35B-A3B-Preview

Explore Clover-35B-A3B-Preview from DiffuSpace: published specifications, source-linked vendor benchmarks, independent evaluator coverage and recorded pricing when available.

Published specifications

Provider: DiffuSpace
Access: Open weights
License: Apache-2.0
Context window: 128K
Total parameters: 35B
Active parameters: 3B
Released: 2026-10-28
Modalities: text
Family: Clover

Hugging Face · Announcement

Model notes

Open-weight block-diffusion language model (dLLM), DiffuSpace's first release; text-only, thinking mode supported. Derived from Qwen3.6-35B-A3B (hybrid Gated DeltaNet/full-attention MoE, 40 layers, 256 routed experts, top-8 routing, one shared expert) with bidirectional block attention and LowConfidence block decoding (block 32, threshold 0.95). Official ungated Apache-2.0 weights; single commit 2026-10-28 03:07 UTC. Card states 35B total / approximately 3B active and a 131,072-token (128K) context including output; Hugging Face safetensors metadata counts 34,661,350,848 BF16 parameters. Also published on ModelScope and GitHub with SGLang and llama.cpp implementations. Checked 28 October: no OpenRouter route and no Artificial Analysis, Vals AI, ARC Prize, Andon Labs or Bug Hunt Bench entry. Vendor speed figures are local hardware observations, not API speeds.

Compare Clover-35B-A3B-Preview with other models →Compare subscription plans Explore data coverage

Pricing · OpenRouter

No OpenRouter pricing is recorded for this model. Missing rates are not free.

Official / vendor benchmarks

Default headline records. Own-vendor, peer-vendor and third-party provenance remain visible in evidence; configurations may differ.

Official / vendor headline scores; expand evidence for every record
Benchmark / evaluator	Headline score	Evidence
GPQA Diamond	83.8%unknown effort	All 1 recorded result & sources 83.8% · raw 83.8 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "GPQA Diamond". EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
Humanity's Last Exam text-only (DiffuSpace EvalScope; tool use unstated)	12.8%unknown effort	All 1 recorded result & sources 12.8% · raw 12.8 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "HLE text-only". EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
MMLU-Pro	84.8%unknown effort	All 1 recorded result & sources 84.8% · raw 84.8 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "MMLU-Pro". EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
MATH-500	96.2%unknown effort	All 1 recorded result & sources 96.2% · raw 96.2 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "MATH-500". EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
AIME 2026	83.3%unknown effort	All 1 recorded result & sources 83.3% · raw 83.3 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "AIME 2026 (avg@8)". avg@8. EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
LiveCodeBench v6	72.6%unknown effort	All 1 recorded result & sources 72.6% · raw 72.6 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "LiveCodeBench v6". EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
SciCode (DiffuSpace EvalScope; aggregation unspecified)	30.9%unknown effort	All 1 recorded result & sources 30.9% · raw 30.9 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "SciCode". EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
ZebraLogic	95.8%unknown effort	All 1 recorded result & sources 95.8% · raw 95.8 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "ZebraLogic". EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
IFBench	72.3%unknown effort	All 1 recorded result & sources 72.3% · raw 72.3 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "IFBench". IFBench variant (strict/loose) not stated. EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
AA-LCR	72%unknown effort	All 1 recorded result & sources 72% · raw 72 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "AA-LCR". Vendor EvalScope run, not an Artificial Analysis harness observation; revision not stated. EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
τ³-bench (overall)	65.3%unknown effort	All 1 recorded result & sources 65.3% · raw 65.3 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "τ³ Overall". EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
BFCL v4 (overall)	67.9%unknown effort	All 1 recorded result & sources 67.9% · raw 67.9 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "BFCL-v4". Overall BFCL-v4 as labeled. EvalScope; DiffuSpace reproduced every baseline under the same settings with each model's recommended decoding. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
GDPval-AA v2 Elo (DiffuSpace local NeMo Gym run)	1091unknown effort	All 1 recorded result & sources 1091 · raw 1091 Elo Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "GDPval-AA v2 (Elo)". Evaluated locally with the NVIDIA NeMo Gym GDPval-AA v2 recipe. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
Terminal-Bench 2.1	51.7%unknown effort	All 1 recorded result & sources 51.7% · raw 51.7 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "Terminal-Bench 2.1". Harness: Terminus 2, each task run three times and averaged. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
Terminal-Bench Core v0.1.1 hard subset (21 tasks, DiffuSpace)	38.1%unknown effort	All 1 recorded result & sources 38.1% · raw 38.1 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "Terminal-Hard". 21 hard tasks from Terminal-Bench Core v0.1.1; Terminus 2, each task run three times and averaged. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
SWE-bench Verified	73.6%unknown effort	All 1 recorded result & sources 73.6% · raw 73.6 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "SWE-bench Verified". Harness: Codex, each task run three times and averaged. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
SWE-Pro (DiffuSpace label; Codex harness, revision unspecified)	46.5%unknown effort	All 1 recorded result & sources 46.5% · raw 46.5 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog general-benchmarks.jpg table (same values as the agentic-benchmarks.jpg chart and the 28 October X post): "SWE-Pro". Harness: Codex, each task run three times and averaged. Thinking mode; effort label not published. https://diffuspace.ai/blog/clover-35b/
τ³-bench Banking	14.4%unknown effort	All 1 recorded result & sources 14.4% · raw 14.4 % Headline · unknown effort · Own vendor Source/record date: 2026-10-28 From the DiffuSpace launch blog on-device-benchmarks.jpg table: "τ³ Banking". In-house evaluation. https://diffuspace.ai/blog/clover-35b/
Clover GB300 · internal agent workloads · overall generation speed	680.0 tokens/sunknown effort	All 1 recorded result & sources 680.0 tokens/s · raw 680 tokens/s Headline · unknown effort · Own vendor Source/record date: 2026-10-28 Launch blog text; local vendor hardware observation, not an API speed. https://diffuspace.ai/blog/clover-35b/
Clover H200 · agent tasks · generation speed	515.0 tokens/sunknown effort	All 1 recorded result & sources 515.0 tokens/s · raw 515 tokens/s Headline · unknown effort · Own vendor Source/record date: 2026-10-28 Launch blog text; local vendor hardware observation, not an API speed. https://diffuspace.ai/blog/clover-35b/
Clover RTX 5070 Ti · INT4 · MATH-500 generation speed	101.8 tokens/sunknown effort	All 1 recorded result & sources 101.8 tokens/s · raw 101.8 tokens/s Headline · unknown effort · Own vendor Source/record date: 2026-10-28 Launch blog text; local vendor hardware observation, not an API speed. https://diffuspace.ai/blog/clover-35b/

Independent evaluators

Evaluator harnesses are distinct from vendor measurements. Missing coverage is not a failed test.

No independent evaluator results are recorded for this model.

Read how we select and source scores or the comparison guide.