Diffusion Intelligence releases Clover-35B-A3B-Preview model
The story so far
Diffusion Intelligence has released the Clover-35B-A3B-Preview model with 3.5 billion parameters, claiming it is twice as fast as the Jev+Mercury 2.5 combination. In the WindTunnel web agent evaluation, its median execution time is 1.52 seconds with a success rate of 98.6%. The model adopts a MoE architecture, supports thinking mode, and achieved a score of 73.6 on SWE-bench Verified. Although the name includes 'Diffusion', the actual experience is similar to familiar large language models in terms of usage, whether writing code, operating web pages, calling tools, or generating HTML. For example, to generate a solar system simulator, Clover thinks for about 1 second, and then it takes 4-5 seconds to get an interactive solar system simulator. Another example is to give a pile of resume information and let it organize it into LaTeX format resume code, which takes about ten seconds to complete. In the WindTunnel web agent evaluation of 49 tasks, each task is tried three times, Clover's median execution time is 1.52 seconds, while the Jev+Mercury 2.5 combination is 3.21 seconds, half more time saved. Diffusion Intelligence DiffuSpace has also released the open-source version of the model under the Apache license.
AI-written from the reports · Updated 20 days ago
Coverage timeline
Follow the coverage to see every side of the story.
- ITHome · AIDiffuSpace releases open-source diffusion language model Clover-35B-A3B-Preview
DiffuSpace has released an open-source diffusion language model Clover-35B-A3B-Preview, which adopts a MoE architecture and outperforms same-scale autoregressive models like Qwen3.6-35B-A3B in multiple benchmark tests, particularly excelling in agent programming and tool use tasks.
- QbitAIDiffusion Intelligence releases Clover-35B-A3B-Preview, a 350 billion parameter model that is twice as fast as Jev.
Diffusion Intelligence releases Clover-35B-A3B-Preview model, with 350 billion parameters achieving a median execution time of 1.52 seconds in the web Agent evaluation WindTunnel, twice as fast as the Jev+Mercury 2.5 combination, with a success rate of 98.6%. The model adopts a MoE architecture, supports thinking mode, and scores 73.6 on SWE-bench Verified.