AI coding tools
Clover-35B-A3B-Preview: The Dream 7B Team's Diffusion LLM Outruns Mercury 2.5 on Agent Benchmarks (Apache-2.0, Team-Reported)
Verdict: DiffuSpace — a startup spun out of the HKU NLP lab behind Dream 7B — released Clover-35B-A3B-Preview on Oct 27 under Apache-2.0: a 35B-total/3B-active MoE built on Alibaba's Qwen3.6-35B-A3B backbone, retrained for block-diffusion generation. The team reports it beating Mercury 2.5 and same-size autoregressive Qwen models on agent and coding benchmarks (SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7), and running at 515 tokens/s on an H200 where a DFlash2-accelerated Qwen baseline does 365. Every benchmark is the team's own selection, none of it independently replicated yet, and "Preview" means preview. But the direction is real: diffusion LLMs are graduating from demo novelties to agent-workgrade tools, and this is the first open-weights entrant big enough to test yourself. If you run latency-sensitive agent loops, it's worth an afternoon; it is not yet a drop-in replacement for your AR stack.
Last checked 2026-10-29. Weights released 2026-10-27 — two days old at time of writing, 92 stars on GitHub, effectively zero English coverage. All benchmark numbers below are team-reported unless noted.
The pitch is simple and worth taking seriously: autoregressive models generate left-to-right, one token at a time; diffusion LLMs update many positions in parallel and refine over several denoising rounds. Clover uses block diffusion — blocks advance sequentially, but positions inside a block update in parallel — so one forward pass can push several tokens toward completion at once. The architecture keeps a causal prefix (your prompt stays causal) while using bidirectional attention inside each generation block, with added Gated DeltaNet parameters to support the diffusion process. Context window: 128K.
Why this matters for coding agents specifically: the historical weakness of diffusion LLMs in agentic work was instability — malformed tool calls, retry loops, failure to react to tool output. Clover's headline demos are precisely there: a 49-task web-agent suite (WindTunnel) where it completed each task 3 times with a median execution time of 1.52 seconds, versus 8.86 seconds for Gemini 3.6 Flash in the same setup, and — per QbitAI's reading of the team's data — 98.6% attempt success versus 95.9% for a Jev + Mercury 2.5 combo that needed a median 3.21 seconds. On a private-project setup task it finished in 13 steps/169.9 seconds where Qwen3.6 Flash took 21 steps/392.3 seconds.
The benchmark table from the launch blog (team-reported, self-selected comparison set):
- SWE-bench Verified: 73.6 — +28.0 over MiniCPM-V 4.7, the other 35B-A3B in the table
- Terminal-Bench 2.1: 51.7 (+15.7 over the same)
- τ³ (tau3) Overall: 65.3; BFCL-v4: 67.9 — claimed best among listed models including Mercury 2.5 and LLaDA2.2 flash
- General knowledge holds up: GPQA Diamond 83.8, MMLU-Pro 84.8, ZebraLogic 95.8, AA-LCR 72.0
Speed, on hardware you can name: single H200, agent workloads, 515 tokens/s versus 365 for the same-size Qwen model with DFlash2 acceleration; single GB300 at BF16, batch size 1, 680 tokens/s on internal agent data. Weights are roughly 69GB in BF16 — that's an 80GB-class card for the full model, with ~3B parameters active per token doing the compute. Serving is supported on SGLang and llama.cpp, with a Transformers quick-start path for testing.
What to hold at arm's length: every comparison above comes from DiffuSpace's own launch blog, against a comparison set the team chose. WindTunnel is not an industry-standard benchmark. Mercury 2.5 is closed-weight, so "beats Mercury" rests entirely on the team's harness. The 98.6%-vs-95.9% Jev+Mercury comparison is transcribed from QbitAI's report, not the blog. And the model is a Preview — first release, 92 GitHub stars, no community evals yet. The honest read: the direction is credible, the specific numbers are unverified, and the only benchmark that matters for you is your own workload.
Who should actually spend time on this: developers building latency-sensitive agent loops — browser automation, form-filling, multi-step tool chains — where 1.5-second versus 8-second step latency compounds into a different product. Who shouldn't: anyone needing a drop-in replacement for an established AR stack today, anyone below 80GB VRAM who can't quantize, and anyone who needs battle-tested tool-calling reliability without a validation phase.
The meta-story is arguably bigger than the model: Dream 7B (April 2025, HKU NLP, 2.7M+ downloads) proved diffusion LMs could spar with autoregressive ones at 7B. Clover answers the next question — can the paradigm hold at 35B MoE scale — and the answer, per team data, is yes while keeping the speed advantage. The other quiet detail: the backbone is Alibaba's Qwen3.6-35B-A3B. China's open-model stack is now compounding — a Hong Kong lab's research lineage trains on a mainland lab's open backbone and ships under Apache-2.0 for anyone to fork. And DiffuSpace has already teased what's next: a decision-model line in the vein of TypeSafe's Jev, using the diffusion trick to answer up to 8 related questions in a single denoising pass — internally validated at 95.0% accuracy versus 89.8% answering independently.
Related
- Decision model pricing: Microsoft Decision-1 — the market DiffuSpace wants to enter next
- Ant's LingDT decision model — China's other bet that structured decisions deserve their own model
- MOSS-Image local deployment — why the frozen backbone, not the headline number, sets your VRAM bill
Sources: DiffuSpace/Clover-35B-A3B-Preview repo (README: architecture, Qwen3.6-35B-A3B backbone, 69GB BF16, 128K context, backends; created 2026-10-27, 92 stars via GitHub API) and the official launch blog (benchmark table, WindTunnel runs, GB300/H200 throughput, Dream 7B lineage, joint-mode decision-model roadmap — all team-reported); QbitAI 2026-10-29 report (Jev+Mercury comparison: 3.21s median / 95.9% success — transcribed, not independently verified).