Axios reports Reflection AI is about to release an open‑weight model
Axios says Reflection AI's first open-weight model is close and should compete with top Chinese open models. Reflection, which pays SpaceX $150 million a month for compute, declined to comment.
Reflection AI is about to release an open-weight model that should compete with the leading Chinese open-weight models, Axios reported on Sunday, citing unnamed sources. The startup, which is backed by Nvidia, has not announced a release. Reflection declined to comment to Axios, according to a summary of the report.
The model is expected to lag the most capable U.S. models at first, Axios said, but to be strong enough for companies to build their own low-cost proprietary systems on it. Axios also reported that other Western labs plan open-weight releases this month. It did not name them in the account the Post could read.
https://x.com/AndrewCurran_/status/2106846163534241912
The report gives no model name, parameter count, licence, benchmark or date. Nobody outside Reflection has tested the model, and the Post has not seen weights, a model card or a Hugging Face page. Axios's full text was not available to this paper, so the details above come from its headline and published summaries.
The compute behind it is better documented. TechCrunch reported in June that Reflection pays SpaceX about $150 million a month for Nvidia GB300 chips at the Colossus 2 site near Memphis. Andrew Curran put a separate Nebius deal at $1 billion. He said that makes a strong model plausible.
Reflection has slipped its own timeline before. It first targeted early 2026 and later said late this year, and forecasters had pointed to January 2027. Some open-model watchers are sceptical: the poster teortaxesTex wrote that training a model as good as Z.ai's GLM 5.3 normally takes years, though he allowed that a lull in Chinese releases could let Reflection pull it off.