Reflection AI to release open-weight model aimed at DeepSeek and Qwen

Reflection AI is preparing to release an open-weight foundation model meant to compete with DeepSeek and Alibaba’s Qwen, the Chinese models that have spent the past stretch running laps around Western rivals in the open-source world.

According to an Axios report dated October 4, 2026, the model is expected to arrive this month. It is pitched as a match for the best Chinese open-weight systems.

What Reflection is building

An open-weight model is one whose trained parameters are published, so outside developers can download and run it on their own hardware.

Per the Axios report, Reflection’s model is expected to stand toe-to-toe with the top Chinese offerings. It is also expected to initially trail the leading closed models from OpenAI, Anthropic and Google.

CEO Misha Laskin has described frontier models as being “kind of like rocket ships,” a nod to how much time it takes before they reach full capability.

The money and the machines

Reflection was founded in March 2024 by Laskin and Ioannis Antonoglou, both former researchers at Google DeepMind.

The startup raised $2B at an $8B valuation in October 2025. Subsequent rounds have reportedly pushed that figure as high as more than $25 billion, with reported current valuations ranging between $8B and $25B.

Reflection has struck a compute deal with Nebius worth more than $1 billion. It has also secured substantial capacity from SpaceX’s Colossus, including significant rentals of Nvidia servers.

The company already has one product in the wild. Asimov is a code-comprehension agent designed to help developers make sense of codebases.

The AI factory pitch

Reflection’s business strategy rests on what it calls an “AI factory” model. The idea is to combine its open weights with a client’s proprietary data and advanced Nvidia GPU compute. The intended result is customized, localized AI deployments that cost less to run.

Why Chinese labs set the bar

Chinese laboratories currently sit atop many open-weight model leaderboards. Their models frequently rank ahead in coding, mathematics, reasoning and cost-efficiency. DeepSeek and Qwen have become reference points for what open models can do.

What this means

For enterprises, the stakes are practical. If Reflection’s model genuinely matches the top Chinese open-weight systems, buyers gain another credible option that they can host, customize and control.

For investors, the valuation leaves little room for a soft debut. A company reportedly valued at up to more than $25 billion will be judged on benchmarks and adoption, not on press coverage.

Reflection is explicitly expecting to trail OpenAI, Anthropic and Google at first, betting that later iterations close the gap.

The things to watch are straightforward. They include independent benchmark results against DeepSeek and Qwen, the licensing terms attached to the weights, and which enterprises sign up for the factory model first.