Startup profile · funding coverage
DiffuSpace funding, valuation and investors
Chinese diffusion-language-model startup behind open-weight Clover-35B-A3B-Preview and the Dream research lineage.
Verified company research · funding, investors and latest developments
Keep track of DiffuSpace
Save this profile to your VCT watchlist for a quick return.
Funding, valuation & investors
Answer-first snapshotLatest funding
Two rounds announced together (tranche dates and amounts undisclosed)
Several hundred million CNY in aggregate; near CN¥500M per unnamed sources · October 9, 2026
Latest known valuation
Not publicly disclosed
Total disclosed equity funding
Not publicly disclosed
Excludes debt, grants, acquisitions, secondaries, and IPO proceeds.
Current status
private
Stage not recorded · Shenzhen, Guangdong, China
Investors in latest funding
Lead: Matrix Partners China , Shunwei Capital , Legend Capital
Other: CAS Star , Huawei Hubble Investment , Horizon Robotics
Overview
DiffuSpace was founded in Shenzhen in May 2026 by researchers from the University of Hong Kong NLP group associated with Dream 7B. Chinese media reported on October 9 that the company had completed two financing rounds co-led by Matrix Partners China, Shunwei Capital and Legend Capital, with aggregate funding near CN¥500M according to anonymous investor sources; no exact figures were company-disclosed.
On October 30 it released the open-weight Clover-35B-A3B-Preview, a block-diffusion 35B-total/3B-active MoE language model. The team reports 680 tokens/s on a single NVIDIA GB300 on its own agent workload, not independently established throughput across tasks.
Why DiffuSpace is interesting
A direct diffusion-text-generation competitor to Inception and Zyphra. Following its October 2026 financing disclosures, DiffuSpace released Clover-35B-A3B-Preview with downloadable Apache-licensed model weights. Company benchmarks show promising speed and coding performance but need outside replication.
Product & use cases
Clover-35B-A3B-Preview, an Apache-licensed open-weight block diffusion language model with 35B total / 3B active parameters, thinking mode and code/tool-use support. Weights and code on Hugging Face/GitHub, with SGLang and llama.cpp support; earlier Dream 7B work predates DiffuSpace's formation.
Key facts
- Two financing rounds disclosed October 9, 2026; exact company-confirmed amount undisclosed
- Core investors: Matrix Partners China, Shunwei Capital and Legend Capital
- Clover-35B-A3B-Preview released October 30, 2026; open weights, 35B total / 3B active
- Company reports 73.6 SWE-bench Verified and 680 tokens/s on GB300; independent verification pending
Funding history (newest first)
Two rounds announced together (tranche dates and amounts undisclosed)
2026-10-09 Several hundred million CNY in aggregate; near CN¥500M per unnamed sources- Matrix Partners China (lead)
- Shunwei Capital (lead)
- Legend Capital (lead)
- CAS Star (participant)
- Huawei Hubble Investment (participant)
- Horizon Robotics (participant)
Competitive landscape
Edge: Unlike a pure research proposal, Clover offers public weights for third-party testing. The company's reported 680 tokens/s is specific to one GB300, BF16, batch-one agent workload and cannot be compared directly with models on different GPUs.
The October 30, 2026 Clover release creates a more testable product milestone than previously disclosed. The company's 73.6 SWE-bench Verified and 680 tokens/s claims remain self-reported. The company is developing block diffusion, not Stanford's NCM; the two October financing tranches remain undisclosed separately.
-
Inception Labs direct
Commercial US diffusion-model company.
-
Google DeepMind incumbent
Gemini Diffusion research program.
Market / IPO context
Editorial / static context — not a live quote.
FAQs about DiffuSpace
Practical answers founders, operators, and investors typically search for.
Last updated:
Editorial note: AI tools assisted with research, structure, or drafting. Venture Capital Tracker retains human editorial responsibility for factual accuracy, relevance, and source quality before publication.