Nvidia-backed Reflection is preparing a US open-weight model for this month

Reflection, a US startup backed by Nvidia, is preparing to release its first open-weight model within October, Axios reported on October 4. The company was founded in 2024 by two former Google DeepMind researchers, and it raised $2 billion at an $8 billion valuation a year ago. Its first model is expected to trail the top US frontier models at launch. It should be competitive with the leading Chinese open-weight systems, and capable enough for a company to layer its own data on top and build a cheaper custom AI. A Reflection spokesperson declined to comment. Axios · Yahoo Finance

Reflection is not alone. Other Western open-weight models ship this month too, according to Axios. The strongest open-weight systems today are almost all Chinese; the big US labs ship closed models only. That gap is the opening. Nvidia and Reflection want to pair GPUs with open weights, so a company keeps its own data. Reflection has briefed Washington officials in recent weeks on the release and on its longer-term “AI factory” plan, a vision Jensen Huang has pushed for years. The News

A downloadable alternative to the closed APIs is arriving

Until now, near-frontier quality meant the Anthropic, OpenAI, or Google API. A US open-weight model adds a real choice: download the weights, run them yourself, fine-tune on your own data, and leave when a vendor raises prices. Axios reported that combining a cheap open model with good company data can match expensive frontier systems in some situations. Nothing here says to migrate the product you already run. It does add a question to stack decisions: can I run this myself later?

Nothing is confirmed yet. Reflection declined to comment. The model’s performance, license, and price are all unpublished.

  • Worth doing now: keep prompts and eval sets separate from any single model, so a new open-weight release takes an afternoon to A/B rather than a rewrite.
  • Worth watching: the license terms (commercial use, fine-tuning) and the real hardware requirements of this month’s Western open-weight releases.

The rest of today’s news

  • Trump taps Jay Clayton to lead a “Super Intelligence Force”: the director of national intelligence will chair a task force with the FTC chair and two other officials, reporting within 120 days. CBS News · The Guardian
  • Musk says SpaceXAI will be renamed SpaceXSI: he confirmed the change after Trump’s order replacing “artificial intelligence” with “super intelligence.” Reuters
  • Meta’s Muse instructions say a user’s household authority overrides its safety training: researchers extracted the system prompt; Meta says the files were meant to be user-accessible. Wired
  • Insilico’s AI designs a blood-cancer PROTAC: its Abu Dhabi lab designed BTK-degrading molecules, with human trials still years away. The National
  • YouTube creators launch #TeamHuman: Mark Rober and Kurzgesagt, with Hinton and Bengio signing on, are petitioning for a global AI slowdown. TeamHuman