Reflection AI, Valued at $25 Billion, Prepares to Ship Its First Open-Weight Model, Which Is Expected to Be Slightly Worse Than the Closed Ones
Nvidia-backed Reflection is reportedly about to release an open model pitched as a Western alternative to DeepSeek and Qwen, backed by $7 billion in committed compute and a single shipped product.
Reflection AI has raised, per Startup Fortune, an $800 million Nvidia-led investment as part of a $2 billion Series B, reached a $25 billion pre-money valuation by April, and secured more than $7 billion in committed compute. It has, as of this week, exactly one shipped product: Asimov, a code-comprehension agent. Axios reports the company’s first open-weight model is about to arrive.
The sourcing here is worth stating precisely. Per Axios, as relayed by AI Weekly and ExplainX, Reflection’s first open-weight model is “imminent” and expected this month. ExplainX is explicit that as of October 5 no weights, model name, benchmarks or license have been published. So the model does not exist yet in any way you can download, and this article is a preview of a product, not a review of one.
What is known is the pitch. Per AI Weekly, CEO Misha Laskin wants Western enterprises to have a bank-approved base on which to build proprietary, low-cost AI, positioned as an alternative to DeepSeek and Qwen, whose security profile makes some enterprises nervous. It leans on Nvidia’s “AI factory” idea: hardware plus open weights, so customers keep ownership of their data. Startup Fortune notes that Jensen Huang backs the model of enterprises buying Nvidia chips paired with customizable open weights. Nvidia selling chips to run models that run on Nvidia chips is not a conflict of interest; it is a supply chain.
The honest part
Credit where due: the reporting is refreshingly candid about expectations. ExplainX and Startup Fortune both say the model is expected to “trail the top closed systems” from OpenAI, Anthropic and Google while competing against the Chinese open models. This is a rare pre-launch message: please do not expect the best thing. Most launches arrive with benchmark charts in which the new model wins at everything except, discreetly, the hard parts. Here the expectation is set in advance at “second tier, but ours.”
For enterprises, that might be a perfectly good sales pitch. A model that’s a bit behind the frontier but runs on your own hardware, on your own data, with no foreign-origin questions, is a product category with real customers. It’s the AI equivalent of a reliable sedan.
The compute bill
The money details are the stranger part. Startup Fortune reports the $7 billion in commitments includes a Nebius agreement worth more than $1 billion through 2029, and a SpaceX contract that began July 1, 2026, at $150 million a month, potentially totaling $6.3 billion if fully executed. The SpaceX deal grants access to Nvidia servers, per AI Weekly’s description of the Nebius and SpaceX agreements.
At $150 million a month, Reflection’s compute bill is roughly what a small nation’s ministry of transport would call a “significant line item.” It’s a lot of rented GPUs in service of a model whose weights will then be given away. This is the open-weight business in miniature: spend a fortune to produce a thing, publish it for free, and trust that the real revenue is in everything standing near it. The consultants, the hardware, the support contracts, the “AI factory.”
Who founded it
Reflection was founded, per Startup Fortune, by Misha Laskin, a former Gemini reward-modeling lead at DeepMind, and Ioannis Antonoglou, a co-creator of AlphaGo. Investors in the Series B included Eric Schmidt, Citigroup, 1789 Capital, Lightspeed and Sequoia alongside Nvidia. The valuation tripled from $8 billion in October 2025, a trajectory that will, if nothing else, give the eventual model a lot to live up to.
AI Weekly adds that more Western open-weight launches are anticipated this month, and quotes one commenter speculating that another lab with “rather large amounts of compute” may drop a strong open model soon. So the vibe for October is: a bunch of Western labs, simultaneously, trying to out-open the Chinese labs that had been out-opening everyone.
There’s a satisfying irony in a US-based startup describing its main selling point as “the open-weight model that isn’t from there.” The strategy is less “we have a better model” than “we have a model with better paperwork.” As it turns out, in enterprise software, paperwork often wins.
We’ll find out how the model performs when it appears. Until then: $25 billion, $7 billion in compute, one product, and a launch described by its own press as “probably not the best.” It’s the most transparent possible way to ask for the world’s attention.
Sources: