Reportedly: DiffuSpace raises hundreds of millions of yuan in record dLLM funding
Reported by Zhidongxi · not yet confirmed by the company or a second independent outlet. We update this page when it is.
DiffuSpace, a Shenzhen large model company, has completed two funding rounds totaling hundreds of millions of yuan, setting a record for global diffusion language model (dLLM) financing, Zhidongxi reported. The funding was led by Matrix Partners China, Shunwei Capital, and Legend Capital, with participation from Huawei Hubble and Horizon Robotics. The company plans to release and open-source a new larger dLLM in October.
China context
- Original name
- 扩散智能（DiffuSpace）
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can watch for DiffuSpace's promised October open-source release on Hugging Face to evaluate dLLM performance for on-device agent use cases.
- For investors
- Investors outside China can monitor whether DiffuSpace's dLLM models gain adoption beyond China, as the technology could disrupt autoregressive model dominance in edge AI.
Translated from Chinese. Quotes and facts link to the original sources.
DiffuSpace, founded in 2026 in Shenzhen by Professor Kong Lingpeng and his PhD students, focuses on diffusion language models (dLLMs) as a next-generation foundation model. Unlike autoregressive models, dLLMs use global parallel diffusion generation, making them suitable for on-device agents. The company's Dream 7B model, released in 2025, surpassed autoregressive models of the same size and matched DeepSeek-V3's planning ability, with over 2.5 million downloads on Hugging Face. The new funding will be used for model training, vertical scenario adaptation, and infrastructure development.
DiffuSpace's dLLM approach uses global parallel diffusion generation instead of sequential token generation, which can improve inference speed for on-device agents. The company claims its Dream 7B model outperformed same-size autoregressive models and matched DeepSeek-V3 in planning. A partnership with Acrab reportedly enables 5x faster on-device agent execution.
DiffuSpace's record dLLM funding round, with backing from Huawei Hubble and Horizon Robotics, strengthens competition for on-device AI agents by providing capital for larger model training and open-source releases. This pressures other Chinese AI labs to accelerate diffusion language model development or risk losing talent and investor attention in the emerging dLLM segment.
DiffuSpace's dLLM technology targets on-device AI agents, offering potential cost and latency advantages over cloud-based autoregressive models. The funding enables the company to scale model training and pursue vertical applications in AI PCs, smart vehicles, robotics, and smart homes.
The next verifiable signal is DiffuSpace's planned October release and open-sourcing of a larger dLLM. If the model is released on Hugging Face with weights, it will confirm the company's open-source commitment and allow independent benchmarking against autoregressive models.