Reflection AI's First Open-Weight Model Is Coming: Taking On DeepSeek and Qwen

The source: Axios’s October 4 exclusive

Axios reported on October 4, local time, that the Nvidia-backed AI startup Reflection AI is preparing to release its first open-weight foundation model, expected this month. The report cites unnamed sources; Reflection declined to comment.

One calibration first: the model hasn’t been released yet. No model name, parameter count, license, benchmark scores or launch date have been announced. Some aggregators ran headlines saying it was “released” — that’s an exaggeration.

What kind of company is Reflection

Reflection was founded in March 2024 by Misha Laskin and Ioannis Antonoglou, both researchers out of Google DeepMind — Antonoglou worked on AlphaGo and Gemini.

In October 2025, the company closed a $2 billion funding round at an $8 billion post-money valuation. Investors include Nvidia, Lightspeed and Sequoia. Later media reports pushed the figure to around $25 billion, but that number comes from media coverage, not from the company.

Its first product is called Asimov, a code-understanding agent that helps developers make sense of codebases, sold for private enterprise deployment.

Burning cash on compute: over $7 billion through 2029

Reflection signed a compute agreement with Nebius worth more than $1 billion, and leased SpaceX’s Colossus cluster. According to earlier reporting from CNBC and Bloomberg, the SpaceX terms run about $150 million a month, paid from July 2026 through 2029 for a total of about $6.3 billion, with the option to exit on 90 days’ notice after the first three months. Axios puts the company’s compute commitments through 2029 at more than $7 billion.

CEO Laskin has compared frontier models to “rockets”: they take a long time to fully mature. (Via Crypto Briefing.)

The bet: a “Western DeepSeek”

The money is aimed at a clear gap: today’s top open-weight models come almost entirely from Chinese companies, with DeepSeek and Alibaba’s Qwen leading the open-source community. Axios reports that Reflection’s model is expected to compete head-on with China’s best open models, though it will initially trail the top closed models from OpenAI, Anthropic and Google.

The bigger idea is the “AI factory”: an enterprise or government combines its own data, Reflection’s model and its own compute into an AI system that is cheap and customizable to its own needs. Axios notes that banks and the Pentagon are wary of using Chinese models over security concerns — and Reflection wants to fill that space. In March, the company signed a memorandum of understanding with South Korea’s Shinsegae Group to build a 250-megawatt AI factory.

The report also notes that other Western open-weight model releases could land this month — not just Reflection.

What we know and what we don’t

What we know: the source is Axios, dated October 4; the positioning is an open-weight foundation model aimed at DeepSeek and Qwen; enterprise users will eventually be able to download it, modify it and customize it with private data; the scale of the funding and compute commitments.

What we don’t: the model name, parameter count, license, benchmark scores, release date, or price. Get your evaluation suites and deployment pipelines ready, and start testing the moment the weights land. Don’t build production plans around a model that hasn’t shipped.