Clover-35B-A3B brings diffusion models into agent work
Diffusion Intelligence released Clover-35B-A3B, an open preview that targets coding, tool use and browser agents with faster parallel generation.
A different route to agent speed
Diffusion Intelligence has released Clover-35B-A3B-Preview, a 35-billion-parameter mixture-of-experts diffusion language model designed for coding, webpage generation and tool-using agents. About 3 billion parameters are activated per token. Unlike conventional autoregressive models, Clover uses block diffusion and repeated denoising to update groups of tokens in parallel.
The release matters because the team is testing diffusion language models on practical agent workloads. In demonstrations, Clover generated interactive HTML, handled browser workflows and completed multi-step operations through WebMCP. In the company’s full evaluation of 49 tasks with three attempts per task, Clover’s median agent execution time was 1.52 seconds, compared with 8.86 seconds for Gemini 3.6 Flash. The company also reports that both models completed the six selected WebMCP tasks in the demonstration.
The unresolved engineering question
Clover’s architecture combines block diffusion with a redesigned bidirectional Gated DeltaNet. The company says historical context is compressed into a recurrent state, while tokens within the current block interact through bidirectional attention. Reported scores include 65.3 on τ³ Overall, 67.9 on BFCL-v4, 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1.
Those numbers are company-reported and depend on the stated hardware, prompts, tools and evaluation setup. The model is still labeled a preview. Long-horizon reliability, reproducibility and performance under changing tool outputs are not yet independently established. If Clover’s speed survives broader testing, diffusion models could become a credible alternative for latency-sensitive agents rather than a research curiosity.
Uncle Cat take
Clover’s strongest claim is not 35B parameters; it is finishing six tool steps without losing the thread. That execution stability deserves verification before the speed headline is believed.