ShenzhenA diffusion language model reached 680 tokens a second.
China’s DiffuSpace says its own tests have it beating same-size Qwen on 6 agent and coding benchmarks.
Diffusion models draft blocks of text in parallel and refine them, instead of writing one token at a time.
The speed is on one Nvidia GB300 chip, and the 35-billion-parameter model is a preview.
A companion decision model, scoring 95.0% on an internal test, will follow later.
How each outlet framed it
- Zhidx leans favorable
- foregrounds the joint-mode decision model, where 95.0% accuracy on linked questions beats independent answering, and the Dream 7B lineage
Sources: Zhidx