Skip to content
Source · Daily Brief

AI Daily Brief — 22 September 2025

The week opened with two major Chinese open-source releases hitting on the same day: DeepSeek V3.1-Terminus (671B parameters, MIT) addressed user-reported language-mixing glitches and pushed agentic benchmarks sharply higher, and Alibaba’s Qwen3-Omni shipped a native end-to-end omni-modal MoE across 119 text languages. Compute politics drove the US story: NVIDIA and OpenAI announced a $100B letter of intent to deploy 10 GW of NVIDIA systems progressively, the first gigawatt on Vera Rubin in H2 2026. Distyl AI raised $175M at a $1.8B valuation, and plaintiffs in Bartz v. Anthropic filed their allocation plan ahead of Judge Alsup’s preliminary-approval hearing.

Top stories

  • DeepSeek V3.1-Terminus ships under MIT. 671B params / 37B active / 163.8K context. BrowseComp 30.0 → 38.5, SimpleQA 96.8, SWE-bench Verified 68.4, Terminal-bench 31.3 → 36.7. Marketed as a service-pack stabilizer focused on Chinese/English code-switching cleanup and agentic tool use. via DeepSeek
  • Qwen3-Omni: open-source omni-modal MoE. Alibaba shipped Qwen3-Omni Apache 2.0 — text/image/audio/video in and out, 119 text + 19 speech-in + 10 speech-out languages, 234 ms audio and 547 ms video streaming latency, claiming open-source SOTA on 32 of 36 audio/AV benchmarks. via VentureBeat
  • NVIDIA-OpenAI sign $100B letter of intent. NVIDIA pledged up to $100B to OpenAI deployed progressively as each gigawatt comes online, with the first GW on the Vera Rubin platform in H2 2026 — pre-staging the Stargate five-site expansion that would land Tuesday. via NVIDIA Newsroom
  • Distyl AI raises $175M at $1.8B. Enterprise-AI consultancy (ex-Palantir founders Arjun Prakash and Derek Ho) closed a Lightspeed and Khosla-led round at 9× its prior valuation, with DST Global, Coatue and Dell Technologies Capital in. Claims 100% production deployment record across Fortune 500. via PR Newswire
  • Bartz plaintiffs file allocation plan. Authors proposed a 50-50 split with publishers for most books in the $1.5B Anthropic settlement, with bespoke treatment for education titles — teeing up the Sep 23 preliminary-approval hearing. via Inside Tech Law
  • Oracle Stargate fact sheet hits Capitol Hill. An internal Oracle/OpenAI Stargate fact sheet dated 09-22 was filed with Rep. Arrington’s office, previewing the five new sites (Shackelford TX, Doña Ana NM, Lordstown OH, Milam TX + Midwest) ahead of Tuesday’s formal unveil. via US House

Who shipped

DeepSeek shipped V3.1-Terminus. Alibaba/Qwen shipped Qwen3-Omni under Apache 2.0. NVIDIA and OpenAI announced their $100B partnership. Distyl AI closed Series B. OpenAI, Anthropic, Google DeepMind, Meta stayed quiet on product launches.

Open-source pulse

The biggest open-source day in weeks. DeepSeek V3.1-Terminus and Qwen3-Omni landed within hours, both with permissive licenses and immediate serverless availability (Vercel AI Gateway, SambaNova, OpenRouter, Novita picked up Terminus same-day). Both Chinese — the open-source center of gravity continues to drift east.

Money, infra & hardware

NVIDIA-OpenAI’s $100B LoI dwarfed everything else. Distyl’s $175M at $1.8B was the only other major round disclosed. Oracle’s Capitol Hill fact sheet hinted at the ~7 GW Stargate buildout that would be formally announced the next day.

Quiet corners

No standout arXiv alignment paper dated Sep 22 specifically — the day’s research signal was the DeepSeek V3.1-Terminus model card itself, functioning as a de facto technical release note. No EU AI Act movement, no US Congress hearings beyond the Stargate filing.

By the numbers

  • $100B — NVIDIA’s letter of intent to OpenAI
  • 671B — DeepSeek V3.1-Terminus parameters
  • 119 — text languages Qwen3-Omni handles
  • $1.8B — Distyl AI post-money valuation
  • Most-mentioned company: DeepSeek
  • Quietest segment: US frontier labs

Compiled by AI Feed’s editor from verified web sources for 22 September 2025.