AnswersPublished21 sources
What Clover-35B-A3B Does—and What Its Early Results Don’t Yet Prove
Clover 35B A3B Preview is an open weight diffusion language model built for coding and tool using agents, with 35B total parameters and about 3B active per token. It generates text in blocks with bidirectional attention within each block, rather than strictly one token at a time; that design alone does not prove val...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Clover 35B A3B Preview is an open weight diffusion language model built for coding and tool using agents, with 35B total parameters and about 3B active per token.
- It generates text in blocks with bidirectional attention within each block, rather than strictly one token at a time; that design alone does not prove valid tool calls or reliable recovery from failed steps.
- DiffuSpace reports up to 680 tokens per second on one GB300; a separate web agent report gives a 1.52 second median across 49 tasks.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is DiffuSpace’s Clover-35B-A3B-Preview, open-sourced on October 29, 2026 by the team behind Dream 7B, and how does it target coding, to. Article summary: Clover-35B-A3B-Preview is DiffuSpace’s open-weight, text-only diffusion language model, adapted from Qwen3.6-35B-A3B and associated with the HKU team behind Dream 7B; its reported emphasis is agentic coding and tool use . Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
DiffuSpace’s Clover-35B-A3B-Preview is an open-weight, text-only diffusion language model adapted from Qwen3.6-35B-A3B. Its design and reported evaluations focus on coding and tool-using agents, rather than only conventional chat. DiffuSpace reports strong agent-task scores, but the available evidence does not independently verify those results or establish that the model reliably handles long, error-prone workflows. 3
5
21
How Clover’s block diffusion works
Autoregressive language models typically generate text from left to right, one token after another. Clover instead uses block diffusion: it generates blocks of text with bidirectional attention within each block while preserving a causal prefix—the text established before the current block. 2
4
The model is based on Qwen3.6-35B-A3B and is described as a mixture-of-experts system with 35 billion total parameters and approximately 3 billion active per token. Model descriptions also list a thinking mode and a context window of about 128K tokens. 5
8
Available descriptions mention learned Gated DeltaNet parameters as part of the block-diffusion design. They do not provide enough detail to assess the specific role of any proposed Gated DeltaNet restructuring or learnable block-size calibration. 2
7
What the agent claims cover—and what they don’t
DiffuSpace positions Clover for agentic coding and tool use. Its reported results include software-repair, terminal-task and function-calling benchmarks. A separate report describes a web-agent evaluation and demonstrations involving tasks such as enrollment, invoice handling and information retrieval. 21
The evidence provided does not establish how well Clover performs at front-end generation. Nor does it show that the block-diffusion approach, by itself, prevents malformed tool calls or repeated failed steps. Those are reliability questions that require task-level testing; generation speed or a benchmark score cannot answer them on its own.
Reported benchmark results
The following scores are reported by DiffuSpace or in coverage describing the team’s results; they should not be read as independently verified findings. 21
| Evaluation | Reported score |
|---|---|
| SWE-bench Verified | 73.6 |
| Terminal-Bench 2.1 | 51.7 |
| τ³ Overall | 65.3 |
| BFCL-v4 | 67.9 |
The available sources do not provide enough comparable evaluation detail to establish exact margins over Qwen3.6-35B-A3B or other similarly sized autoregressive models. Treat comparisons cautiously unless the models’ settings and test conditions are aligned.
Speed and web-agent results
A model comparison page attributes a peak of 680 tokens per second on a single NVIDIA GB300 to DiffuSpace’s measurements. 27 In a separate WindTunnel web-agent report covering 49 tasks, each attempted three times, Clover’s median execution time was reported as 1.52 seconds, compared with 3.21 seconds for Jev plus Mercury 2.5. The same report gives attempt success rates of 98.6% and 95.9%, respectively. These are reported evaluation results, not proof of performance on every browser workflow.
21
The available sources do not substantiate a Clover speed figure on a single H200 or a direct speed comparison with DFlash2 speculative decoding. A community ZeroGPU demo is documented, but a hosted demo is not itself a benchmark. 1
Weights, license and release timing
The official Hugging Face model repository hosts Clover’s weights, and DiffuSpace lists an official GGUF repository for local inference. 5
6 PC Watch reports an Apache 2.0 license and says Transformers inference code is available, while support code and setup instructions for SGLang and llama.cpp were described as forthcoming.
17
The release date is reported inconsistently: PC Watch gives October 28 in China time, while DiffuSpace’s Hugging Face launch article is dated October 30. 3
17 The available sources also document a community ZeroGPU chat demo, but do not verify planned agent-harness or desktop-coding-client details.
1
The practical takeaway
Clover is a notable open-weight attempt to bring block diffusion to coding and tool-using agents: its architecture differs from left-to-right token generation, and its reported benchmark and speed results merit follow-up. But those claims remain provisional. Independent tests should examine valid tool-call formatting, recovery after errors, completion of longer workflows and end-to-end latency—not just tokens per second.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What Clover-35B-A3B Does—and What Its Early Results Don’t Yet Prove"?
Clover 35B A3B Preview is an open weight diffusion language model built for coding and tool using agents, with 35B total parameters and about 3B active per token.
What are the key points to validate first?
Clover 35B A3B Preview is an open weight diffusion language model built for coding and tool using agents, with 35B total parameters and about 3B active per token. It generates text in blocks with bidirectional attention within each block, rather than strictly one token at a time; that design alone does not prove valid tool calls or reliable recovery from failed steps.
What should I do next in practice?
DiffuSpace reports up to 680 tokens per second on one GB300; a separate web agent report gives a 1.52 second median across 49 tasks.