Is Clover-35B-A3B-Preview state of the art?Pending benchmarks.Announced Oct 28, 2026; no independent results yet.
DiffuSpace's first model: a text-only diffusion language model that writes in blocks, filling masked positions in parallel instead of one token at a time. It is derived from Qwen3.6-35B-A3B (35B total, about 3B active, mixture of experts), supports a thinking mode and 128K context, and is published on Hugging Face and ModelScope under Apache-2.0.
The details so far
- Lab
- DiffuSpace (team behind HKU NLP's Dream 7B diffusion model, per the launch blog)
- Announced
- Oct 28, 2026
- API
- Not yet
- Open weights
- Published on Hugging Face and ModelScope under Apache-2.0; BF16 weights are about 69 GB
- Parameters
- 35B total, about 3B active (34.66B in the safetensors)
- Type
- Block-diffusion language model (dLLM), text only, with a thinking mode
- Status
- Preview
- Base model
- Qwen/Qwen3.6-35B-A3B
- Architecture
- Hybrid Gated DeltaNet / full-attention MoE, 40 layers, 256 routed experts, top-8 routing plus 1 shared expert; bidirectional attention within each block
- Context
- 131,072 tokens (prompt and response together)
- Serving
- SGLang and llama.cpp implementations in the GitHub repo DiffuSpace/Clover-35B-A3B-Preview, plus a Transformers path
- Hosted access
- Online access by application (chat.diffuspace.ai); no public API announced
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 680 tokens/s overall on internal agent workloads on one NVIDIA GB300 (BF16, batch size 1), and 515 tokens/s on agent tasks on one H200, against 365 tokens/s for a DFlash2-accelerated Qwen setup.
Lab claim: 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, 65.3 on τ³ Overall, 83.8 on GPQA Diamond and 84.8 on MMLU-Pro, ahead of LLaDA2.2 flash, Mercury 2.5, Qwen3.6-35B-A3B and Qwen3.5-35B-A3B on its agent benchmarks.
Lab claim (DiffuSpace on X): the first open-weight production diffusion language model (dLLM).
Sources
LABDiffuSpace announcement on Xx.com/DiffuSpace_Labs/status/2115288124691820738 · Oct 28, 2026LABDiffuSpace blog: Clover-35B-A3B-Previewdiffuspace.ai/blog/clover-35b/REPOClover-35B-A3B-Preview on Hugging Facehuggingface.co/Diffuspace/Clover-35B-A3B-Preview · Oct 28, 2026REPOClover-35B-A3B-Preview on ModelScopemodelscope.cn/models/Diffuspace/Clover-35B-A3B-Preview · Oct 28, 2026REPODiffuSpace/Clover-35B-A3B-Preview on GitHub (code and inference backends)github.com/DiffuSpace/Clover-35B-A3B-PreviewCompare when it lands
Unscored: its base model, Qwen3.6 35B A3B. A scored comparison appears once both are on a shared board.
Tracking continues until an independent board lists it.