DiffuSpace raises hundreds of millions of yuan for diffusion language models
What's the deal? Chinese startup DiffuSpaceDealroom has a profile for this one. Try Dealroom → has closed an early-stage round totalling hundreds of millions of yuan, co-led by Matrix Partners China, Shunwei Capital, and Legend Capital. CAS Star, Huawei's Hubble InvestmentDealroom has a profile for this one. Try Dealroom →, and Horizon Robotics also participated.
The company says the round is the largest ever for a diffusion large language model (dLLM) developer. It will fund model training, vertical-scenario adaptation, and infrastructure research.
What's the endgame? DiffuSpace is building next-generation foundation models based on diffusion rather than the "left-to-right" autoregressive approach used by GPT and Claude. Diffusion models can generate text in parallel and revise any position during generation, which the company says delivers five-to-tenfold inference speedups and stronger planning in some tasks.
It is now training a larger dLLM and plans to release and open-source it soon.
Who's behind it? DiffuSpace was founded in May 2026 in Shenzhen by University of Hong Kong professor Kong Lingpeng and his doctoral students Gong Shansan and Ye Jiacheng. The team's members come from Tsinghua, Peking University, HKU, and Carnegie Mellon.
They have worked on diffusion language models since 2022, when they released DiffuSeq, an early method for applying diffusion to sequence-to-sequence text generation. Their Dream 7B model, released in 2025, surpassed same-size autoregressive models and has passed 2.5 million downloads on Hugging Face.
Why now? Interest in dLLMs is climbing sharply. Research papers in the first eight months of 2026 reached 2.4 times the full-year 2025 total.
GoogleDealroom has a profile for this one. Try Dealroom → has launched Gemini Diffusion, US startup InceptionDealroom has a profile for this one. Try Dealroom → released its Mercury models, and Ant Group has shipped related work. The field is shifting from frontier research toward engineering and application.
Where it's heading: DiffuSpace has partnered with Asian agent-computing platform Acrab, saying the tie-up can speed on-device agents fivefold. The two plan to push dLLMs into AI PCs, smart cars, robots, and smart-home devices.
The signal: A record raise for a diffusion-model startup shows investors are willing to bet on an alternative to the autoregressive architecture that dominates today's large models. With Google, Ant, and Nvidia-watched Inception all active, the question is whether diffusion can scale into a mainstream foundation-model path — or stay a niche speed play.
Read more: sohu.com
Image credit: Generated with Gemini
