Huawei Hubble and Lei Jun's Shunwei Back DiffuSpace: Two Rounds Total Close to 500 Million RMB, a Record for Diffusion Language Models

While the largest foundation-model companies head toward public listings, a Shenzhen startup is betting on a different way to generate text. DiffuSpace, founded in May 2026 by researchers from the University of Hong Kong's NLP lab, has completed two consecutive funding rounds led by Matrix Partners China, Shunwei Capital and Legend Capital, with CAS Star, Huawei's Hubble and Horizon Robotics participating. Chinese business media reported the two rounds total several hundred million yuan - close to 500 million RMB, roughly $70 million, according to people familiar with the matter - which would make it the largest disclosed funding for a diffusion language model (dLLM) startup.

The team is the credential. Co-founder Kong Lingpeng co-directs HKU's NLP lab and previously researched at DeepMind; co-founders Gong Sansan and Ye Jiacheng hold HKU computer science doctorates. The group began exploring diffusion for text in 2022, published DiffuSeq - an ICLR 2023 paper that made diffusion models work on sequence-to-sequence generation - and followed with RDM and DiffuLLaMA. In 2025 the team released Dream 7B, initialized from Qwen2.5 7B and retrained with diffusion; the company says it outperformed same-size autoregressive models and matched the 671-billion-parameter DeepSeek-V3 on planning tasks, and has drawn more than 2.5 million downloads on Hugging Face. Those are company figures.

The technical bet is straightforward. Autoregressive models like GPT generate one token at a time; diffusion models plan a whole answer in parallel and refine it through global iterations, which makes mid-generation correction natural and decoding fast. Chinese coverage of the funding cited up to 10x inference acceleration in some scenarios, and DiffuSpace says a partnership with the agent-computing platform Acrab has shown on-device agents running up to 5x faster on dLLMs. Those performance claims are the company's own and have not been independently verified.

The money is earmarked for model training, infrastructure and adapting the models to vertical scenarios. The company is training a new, larger-parameter dLLM and plans to release and open-source it this month. The investor list hints at where this is heading: Hubble is Huawei's investment arm, Shunwei was co-founded by Xiaomi's Lei Jun, and Horizon Robotics supplies automotive chips - the dLLM pitch is aimed at AI PCs, smart cars, robots and smart-home devices, where low latency and on-device inference matter more than raw scale.

The niche is heating up quickly. Diffusion language models were a curiosity two years ago; Google, Inception and Ant Group have all moved into the area, and Chinese coverage of the round notes that dLLM papers published in the first eight months of 2026 already run at 2.4 times the total for all of 2025. The team has also extended the approach into vision-language work with Dream-VL, vision-language-action with Dream-VLA, and agent variants.

The caveats are real. Diffusion remains a chasing route in language: no dLLM has displaced frontier autoregressive models on quality, and DiffuSpace's benchmark claims come from the company itself. A funding record inside a niche category is not a funding record overall - close to 500 million RMB is a solid early-stage sum in a market where China's top labs now raise in the tens of billions of yuan.

But the round is a data point in a broader shift: as autoregressive scaling gets expensive and crowded, investors are paying for architectural differentiation targeted at the device edge. If the promised open-source release lands this month with independent benchmarks behind it, DiffuSpace will be the first serious test of whether diffusion can move from papers to products.