🚀 RELEASES & LAUNCHES
- DiffuSpace releases Clover-35B-A3B-Preview, an open-weight 35B diffusion language model
DiffuSpace released Clover-35B-A3B-Preview, an open-weight 35B-parameter diffusion language model that generates at 680 tokens per second.- Aimed at developers and agent builders: open weights (Apache 2.0) on Hugging Face, ModelScope, and GitHub, servable via vLLM, SGLang, and llama.cpp.
- Clover is a block-diffusion MoE model with 35B total parameters, ~3B active per token, a 128K context window, and a hybrid Gated DeltaNet/full-attention backbone derived from Qwen3.6-35B-A3B.
- DiffuSpace reports 680 tokens/s on a single NVIDIA GB300 at BF16 (batch size 1) and 515 tokens/s on an H200, claiming 5-10x speedups over same-size autoregressive models.
- Reported benchmarks include SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7, GPQA Diamond 83.8, and MMLU-Pro 84.8, beating LLaDA 2.2 flash, Mercury 2.5, and Qwen3.6-35B-A3B per the company; results are not independently verified.
- It is DiffuSpace's first release, from a lab led by Dream 7B creator Lingpeng Kong, and signals diffusion LLMs becoming competitive at production scale for latency-sensitive agent workflows.