Clover-35B-A3B-Preview

Clover-35B-A3B-Preview

Quick answer: Clover-35B-A3B-Preview is DiffuSpace's first release — a 35B-parameter mixture-of-experts block diffusion language model (~3B active parameters per token) derived from Qwen3.6-35B-A3B's backbone, with a 128K-token context window and a thinking mode. It scores 73.6% on SWE-bench Verified and 51.7% on Terminal-Bench 2.1, leading the diffusion-model comparison set (LLaDA2.2 flash, Mercury 2.5) and similarly sized autoregressive models (Qwen3.6-35B-A3B, Qwen3.5-35B-A3B) on agentic coding tasks. Open weights under Apache 2.0.

At a Glance

Where Clover-35B-A3B-Preview leads

• Highest score among compared models on all six agent benchmarks in DiffuSpace's launch comparison: 73.6% SWE-bench Verified, 51.7% Terminal-Bench 2.1
• Strong knowledge/reasoning: 83.8% GPQA Diamond, 95.8% ZebraLogic, 72.0% AA-LCR
• Faster generation via parallel block diffusion decoding: DiffuSpace reports 680 tok/s on NVIDIA GB300 and 515 tok/s on H200 for internal agent workloads, versus 365 tok/s for a DFlash2-accelerated same-size Qwen model
• Runs INT4-quantized on a 16GB NVIDIA RTX 5070 Ti

Where it lags

• Preview-stage release — DiffuSpace's own comparisons are self-reported, not yet independently reproduced
• Text-only (no vision/audio modality)
• Block-diffusion decoding is architecturally novel; tooling/ecosystem support (vLLM, inference providers) is less mature than for established autoregressive models

Best for: agentic coding and tool-use workloads where decoding speed compounds across long agent loops — DiffuSpace positions it directly against similarly sized autoregressive models on exactly this use case.

What Clover-35B-A3B-Preview Is

Clover-35B-A3B-Preview is DiffuSpace's first model, built on the Hugging Face Transformers implementation of Qwen3.6-35B-A3B's hybrid Gated DeltaNet (GDN) / full-attention MoE backbone (40 layers, 256 routed experts, top-8 routing, one shared expert per layer). DiffuSpace redesigned the GDN into a bidirectional variant for block diffusion: previously generated context is compressed into the GDN's recurrent state, while tokens within the current block interact through bidirectional softmax attention, letting the model learn when to rely on history versus the current block. Generation proceeds via block diffusion with low-confidence decoding — each block starts fully masked and resolves tokens in parallel as confidence crosses a threshold (default block size 32, threshold 0.95).

DiffuSpace's own team traces back to Dream 7B (HKU NLP, April 2025), one of the first diffusion language models shown to compete with autoregressive models; Clover's premise is that the same approach scales to 35B while keeping diffusion's parallel-generation speed advantage. The launch materials emphasize agentic coding and browser/tool-use tasks specifically, citing per-task step counts and wall-clock times (not just benchmark percentages) as evidence that faster generation shortens an agent's full decide-execute-adjust loop.

Specifications

Pricing

Open weights, self-hostable — no per-token API pricing. Available on Hugging Face (~69GB BF16) and ModelScope, with inference support for SGLang and llama.cpp. DiffuSpace reports successful INT4 deployment on a single 16GB NVIDIA RTX 5070 Ti.

Public Benchmark Scores

Scores above are reported by DiffuSpace and shown for context. They are not Benchgen measurements. DiffuSpace's own comparison set includes LLaDA2.2 flash, Mercury 2.5, Qwen3.6-35B-A3B, Qwen3.5-35B-A3B, MiniCPM-V 4.7, LFM2.5-24B-A2B, Nemotron 3.5 Lightning 30B-A3B, and Gemma 4 26B-A4B across different subsets of this table — see the source blog for full per-model methodology.

Excluded from the table above (bucket-C, documented per skill procedure):

• τ³ Overall (65.3) — a generic/unqualified "τ³" score distinct from our domain-specific tau3-banking entity; per standing policy we never silently narrow or widen scope to map a generic name onto a domain-specific entity (same exclusion pattern as prior cycles). The blog separately reports τ³ Banking at 14.4, which maps cleanly and is included above.
• MATH-500 generation speed (101.8 tok/s on RTX 5070 Ti INT4) — a throughput/speed measurement, not an accuracy benchmark score; not a leaderboard-table candidate.
• Internal agent-workload throughput figures (680 tok/s GB300, 515 tok/s H200) and UAV-dispatch / WebMCP / route-planning demo timings — proprietary internal task evaluations with no independently reproducible public benchmark definition.
• Decision model (joint vs. independent answering, 95.0%/80.4% on a 500-group/2,420-question internal validation set) — a separate, not-yet-released DiffuSpace model variant evaluated on an internal dataset; no public spec.

Related Models & Benchmarks

• Compare: Qwen3.6-35B-A3B
• Benchmarks: SWE-bench Verified · Terminal-Bench 2.1 · GPQA Diamond

Frequently Asked Questions

Specs and scores sourced from DiffuSpace's official launch blog (diffuspace.ai/blog/clover-35b) and Hugging Face model card. Last updated 2026-10-31.