DiffuSpace
Clover brings diffusion-style speed to a 35B open language model
DiffuSpace’s preview model drafts text in parallel blocks instead of one token at a time, and reports agent tasks finishing several times faster.
DiffuSpace released Clover-35B-A3B-Preview on 30 October 2026, an open diffusion language model with 35 billion parameters, 3 billion of them active for each token. Members of the team previously built Dream 7B, an earlier diffusion language model released by the University of Hong Kong’s NLP group.
Why diffusion
Most language models write one token at a time, left to right. A diffusion model drafts a block of text and refines it in parallel, which can make generation much faster. Clover uses block diffusion with a redesigned attention scheme: earlier context is compressed into a running state, while tokens inside the current block attend to each other in both directions.
For AI agents, speed is itself a core capability.
Speed and quality, by DiffuSpace’s measurements
- 680 tokens per second on internal agent workloads, on a single NVIDIA GB300 at batch size 1.
- 515 tokens per second on a single H200, against 365 for a similarly sized Qwen model accelerated with DFlash2.
- 101.8 tokens per second on a 16 GB RTX 5070 Ti consumer card with 4-bit quantisation, against 48.5 for similarly sized MiniCPM and Qwen models.
- 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1, ahead of Qwen3.6-35B-A3B and the other diffusion models in its comparison.
- In a web-agent test of 49 tasks, a median execution time of 1.52 seconds, against 8.86 seconds for Gemini 3.6 Flash.
The weights are open under an Apache licence, with inference support in SGLang and llama.cpp. The Hugging Face listing names Qwen3.6-35B-A3B as the base model. DiffuSpace also previewed a decision model that answers several related questions together, which it plans to release later.
What it means for your business
Agent work is a chain of steps: read, decide, call a tool, read again. When every step waits on a slow model, a task that should take seconds takes minutes, and people stop waiting for the agent and do the job themselves. Faster generation at similar quality changes which tasks are worth handing over.
This is a preview with the vendor’s own numbers, so treat it as a signal rather than a recommendation. The signal is that response time is becoming a measure that matters alongside accuracy and price. When you assess any AI tool for your team, time a real task from start to finish; a tool that finishes while the person is still watching is one that gets used.
Sources
- Clover-35B-A3B-Preview: A New Way for Intelligence (opens in a new tab) · DiffuSpace on Hugging Face
- Clover-35B-A3B-Preview model card (opens in a new tab) · DiffuSpace on Hugging Face
General information only, not legal, financial or professional advice. Check the sources for your circumstances.