Clover-35B-A3B-Preview
DiffuSpace open-sourced Clover-35B-A3B-Preview, a 35B MoE block diffusion language model for coding and agents, reporting 73.6 on SWE-bench Verified.
Clover uses block diffusion, where blocks are generated in sequence and positions inside each block are denoised in parallel. The team restructured a Gated DeltaNet backbone for this and added learnable block-size calibration. It is a mixture-of-experts model with about 3B active parameters per token and a thinking mode, and the team reports 515 tokens/s on agent tasks on one H200.
- Date
- Thursday 29 October 2026
- Lab
- DiffuSpace
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | 73.6 | company |
| Terminal-Bench 2.1 | 51.7 | company |
| tau-cubed overall | 65.3 | company |
| BFCL-v4 | 67.9 | company |
| Throughput on agent tasks, one H200 | 515 tokens/s versus 365 tokens/s for a DFlash2 speculative-decoding setup | company |
| Throughput overall, one GB300 | 680 tokens/s | company |
Weights are under Apache 2.0 on Hugging Face, GitHub and ModelScope. All figures are self-reported and unverified, and the model is labeled a preview.
Sources
This record was checked against its sources on 29 October 2026. How we check