AI News
All stories
Reported

DiffuSpace releases open Clover 35B diffusion text model for fast generation

DiffuSpace has released Clover, an open 35B mixture-of-experts diffusion text model with 3B active parameters, which the presenter says runs very fast.

Models · · Updated · 1 min read

DiffuSpace has released Clover, an open-weights text model with 35 billion total parameters and 3 billion active in a mixture-of-experts design. It is built on diffusion. Matthew Berman covered the release in a video roundup of new open models. He said the company claims Clover scores strongly against models of similar size. The benchmark scores appeared on screen, but the transcript gives none of the figures, so the comparison rests on the company's claim.

Speed is the headline claim, and Berman's reasoning for it is architectural. He said that because Clover is a diffusion model, one should expect it to run extremely fast. The video does not explain how the diffusion approach produces text. It shows no throughput measurement, such as tokens per second. The only speed evidence is a single timed task.

That task was a computer-use demo with the prompt "I am at CMU Pittsburgh. How long it takes to drive to the nearest Wendy's." Berman said Clover finished after 169 seconds. Qwen 3.6 Flash, given the same prompt, ran for 578 seconds and did not complete. The transcript says Qwen took 28 steps in total, but it also says the run reached its 30-step limit. The video does not reconcile those two numbers. It is unclear whether Qwen was stopped at the limit or simply used 28 of its steps before stopping.

The comparison is also narrow. Berman showed one prompt and one run per model. The video does not say what hardware either model ran on, what settings were used, or whether the timing was repeated. Those details bear on whether a 169-second versus 578-second result reflects the models or the setup around them.

According to Berman, the weights are available on Hugging Face, and the release is linked from a DiffuSpace preview page on the Hugging Face blog. The model card and the company's scores were not shown in the video in enough detail to check against the on-screen figures.

What changed

DiffuSpace has released Clover, an open-weights text model with 35 billion total parameters, 3 billion active, built on diffusion in a mixture-of-experts design.

Type: New capability · Applies to: Clover

Why it matters

If diffusion text models keep this speed, agent workloads with many steps could become cheaper and faster to run.

What you can do

Teams running multi-step agents could benchmark diffusion models alongside autoregressive ones on latency.

How the story developed

  1. The commentary adds the presenter's timed computer-use demo comparing Clover with Qwen 3.6 Flash.

Original source

We have no link to the original source for this story yet.

Where we saw it

1 source

Forward Future (Matthew Berman)

DiffuSpace

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works