DeepSeek V4.1 Pro enters gray testing as reports say DeepSeek is training a 2T model and planning an 8T one on Huawei chips
Around Sept 28, 2026 Chinese media reported that DeepSeek V4.1 Pro was in gray (staged) testing for some accounts, with default reasoning effort High and a possible release before the Oct 1 holiday. The Information (via TrendForce, Sept 23) reported that DeepSeek is training a ~2T-parameter model and plans an 8T one. Liang Wenfeng reportedly told investors an OpenAI-scale model needs ~50,000 GB300s or ~200,000 Ascend 950s. No official release had been posted as of Sept 29.
Key facts
- Gray test reported Sept 28 by Mydrivers, Sina and Tencent News; default reasoning effort High; most users cannot see it
- No official DeepSeek changelog entry (latest remains V4.1-Flash, Sept 10)
- The Information via TrendForce (Sept 23): a ~2T model in training, an 8T model planned; ~160,000 Ascend 950DT chips planned for an Inner Mongolia datacenter; new Huawei supply expected Q4 2026
- Liang Wenfeng (reported): an OpenAI-scale model needs ~50,000 GB300 or ~200,000 Ascend 950 chips
- DSec paper (arXiv 2609.22978, Sept 19): agentic-training sandbox infra with ~3M sandboxes/day, 380K+ concurrent
- DeepSeek Harness 0.2.0-rc.1 (~Sept 28): first official desktop build
- The Information (Sept 21): Liang Wenfeng called training on Huawei chips one of DeepSeek's biggest bets; Huawei is set to deliver training chips in Q4 2026 or Q1 2027
- The Information (Sept 24): DeepSeek's annualized revenue run rate reached $1B (from under $500M a few months earlier) as it aims to close a ~$7.5B fundraise by late October
- Sept 22: China's CAC reportedly opened a probe into DeepSeek and Moonshot over possible data leaks to Anthropic via Claude (separate entry)
What happened
The next DeepSeek flagship appeared in staged testing while reports described much larger models trained on domestic Huawei silicon.
Why it matters
It shows how far China's top open lab is scaling on non-NVIDIA hardware. Parameter counts and plans are single-source reports (confidence: medium). Update when DeepSeek publishes an official release.
Changelog
- 2026-09-29: created
- 2026-09-29: sweep 2026-09-29: added Huawei-chip bet, $1B revenue run rate and $7.5B raise, and the CAC probe
Related events
- DeepSeek V4.1-Flash: new architecture family, native vision, cheaper API ★★★
- Huawei sets Ascend 950 cluster cloud launch (China Sept 30, global Nov 30) and Ascend 960 roadmap ★★★
- China's cyberspace regulator probes DeepSeek and Moonshot over possible data leaks to Anthropic via Claude ★★★
Sources (7)
- pressMydrivers: DeepSeek V4.1 Pro gray test
- pressSina Tech: DeepSeek V4.1 Pro
- pressTrendForce: DeepSeek reportedly expands Huawei chips to AI training
- paperarXiv 2609.22978: DSec
- codeGitHub: deepseek-harness releases
- pressThe Information: DeepSeek bets big on Huawei chips
- pressThe Information: DeepSeek's annualized revenue hits $1 billion
id: 2026-09-28-deepseek-v4-1-pro-gray-test · updated 2026-09-29 · open in the interactive timeline