What is Clover-35B-A3B-Preview?
Clover-35B-A3B-Preview is the first model from DiffuSpace, a lab whose team came from HKU NLP's Dream 7B diffusion model. Released in late October 2026, it converts Alibaba's Qwen3.6-35B-A3B mixture-of-experts model, 35 billion parameters with about 3 billion active, into a block diffusion model: instead of writing one token at a time, it fills 32-token blocks by resolving the positions it is confident about in parallel.
To do that it adds bidirectional attention inside each block to Qwen's Gated DeltaNet layers, and can restrict each block to a shared subset of experts. It keeps a thinking mode and a 128K-token context, and is text only.
The weights are under Apache 2.0, inherited from Qwen. The Hugging Face code is a reference path; DiffuSpace recommends its own SGLang build for serving, and also provides llama.cpp support.
How much does Clover-35B-A3B-Preview cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| Hosted API | Not offered |
Open weights under Apache 2.0; no hosted API.
Is Clover-35B-A3B-Preview worth it?
Clover's interest is the method: it shows that an existing autoregressive mixture-of-experts model can be converted to parallel block decoding and keep most of its ability. DiffuSpace reports 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, 83.8 on GPQA Diamond and 84.8 on MMLU-Pro, and says it beats the diffusion models LLaDA2.2 flash and Mercury 2.5 and the autoregressive Qwen3.6-35B-A3B on all six agent tests. The comparison numbers for rivals appear only in chart images, and nothing has been measured independently.
The speed case, which is the point of diffusion, is less dramatic than the architecture suggests. On an H200 DiffuSpace measured 515 tokens a second against 365 for Qwen with DFlash2 speculative decoding, about 1.4 times faster. On a 16GB RTX 5070 Ti at INT4 it reports 101.8 tokens a second against 48.5 for similar models. A 680 tokens-a-second figure on a GB300 has no baseline. There is no direct comparison with plain autoregressive decoding of the same model.
The card is candid about rough edges. Long generations can become repetitive, so the defaults include presence and repetition penalties and gradual biases that push the model to end its thinking after 4,096 tokens and its answer after 8,192. The Hugging Face path and DiffuSpace's SGLang kernels can produce different text from the same settings, and BF16 needs about 69GB of GPU memory. The low 14.4 on tau-cubed Banking is a reminder that long multi-turn tasks are where diffusion models still struggle.
For teams watching whether diffusion decoding becomes practical for agents, Clover is the open model to study. For production, the same Qwen3.6 base with speculative decoding gets most of the speed with none of the preview risk.
Where it earns its place
An inference engineer at an AI infrastructure startup
Measure whether block diffusion beats speculative decoding on its own workloads
Serve Clover with DiffuSpace's SGLang build next to Qwen3.6-35B-A3B with DFlash2 on the same H200 and compare tokens a second and task success; DiffuSpace's own gap is about 1.4 times.
A hobbyist developer with a 16GB RTX 5070 Ti
Run a capable reasoning model locally at usable speed
DiffuSpace reports 101.8 tokens a second at INT4 on that card. Expect to tune max_new_tokens, since the defaults push long answers to close after 8,192 tokens.
A university NLP group
Study converting autoregressive models to diffusion
The card documents the added bidirectional Gated DeltaNet parameters, the confidence-threshold decoder (block 32, threshold 0.95) and per-block expert selection, with file-level attribution to Qwen.
Before you commit
- Setup
- Transformers 5.19.0 with trust_remote_code on a GPU with about 69GB free for BF16; for serving, DiffuSpace's SGLang implementation from its GitHub repository.
- Integrations
- Hugging Face Transformers, DiffuSpace's SGLang and llama.cpp forks with self-speculative decoding; Qwen-style chat template with enable_thinking.
- Data and privacy
- Apache 2.0, inherited from Qwen3.6-35B-A3B, with upstream copyright preserved; self-hosted only.
- Cost at scale
- One 80GB H100 or H200 for BF16; INT4 runs on a 16GB consumer GPU per DiffuSpace's tests.
- Limits
- 128K tokens for prompt and output together, including space for the final block; repetition on very long outputs; preview quality; text only.
What’s good
- Reported 73.6 on SWE-bench Verified, well ahead of other open diffusion models on DiffuSpace's tests
- About 1.4 times Qwen with DFlash2 on an H200, per DiffuSpace
- Apache 2.0 with SGLang and llama.cpp implementations
- Keeps a 128K context and thinking mode
What’s not
- Preview; all results self-reported, rivals' scores only in images
- Long outputs can repeat; relies on built-in penalties
- About 69GB in BF16; Transformers and SGLang outputs differ
- Weak 14.4 on tau-cubed Banking
Best for
Researchers and inference engineers evaluating diffusion decoding on a modern MoE model before it reaches production quality.
Clover-35B-A3B-Preview alternatives
Clover-35B-A3B-Preview: common questions
- Is Clover-35B-A3B-Preview worth it?
- Clover is the most capable open diffusion language model yet on its makers' figures, but as a preview with modest speed gains and no outside tests it is for experimenters. We rate it 3.6 out of 5.
- How much does Clover-35B-A3B-Preview cost?
- Open weights under Apache 2.0; no hosted API.
- What are the downsides of Clover-35B-A3B-Preview?
- Preview; all results self-reported, rivals' scores only in images; long outputs can repeat; relies on built-in penalties; about 69GB in BF16; Transformers and SGLang outputs differ; weak 14.4 on tau-cubed Banking.
- Who is Clover-35B-A3B-Preview best for?
- Researchers and inference engineers evaluating diffusion decoding on a modern MoE model before it reaches production quality.
- How much does Clover-35B-A3B-Preview cost for a team or in production?
- One 80GB H100 or H200 for BF16; INT4 runs on a 16GB consumer GPU per DiffuSpace's tests.
- Is Clover-35B-A3B-Preview safe for company data?
- Apache 2.0, inherited from Qwen3.6-35B-A3B, with upstream copyright preserved; self-hosted only.
- What are the best Clover-35B-A3B-Preview alternatives?
- The closest alternatives we have reviewed are Qwen3.8-Flash-Next and Nanbeige4.3-33B.
Answers checked Oct 31, 2026.