Beam: A 501B Open-Weight Model, and Why Nvidia Reportedly Wants Its Maker

Reflection AI released Beam, a 501B-parameter open-weight model for coding and agent tasks that it says needs three to four times less computing than a similar open model. Five days later, the Financial Times reported Nvidia is in talks to buy the startup or raise its stake.
artificial-intelligence
entrepreneurship
Author

Kabui, Charles

Published

2026-10-13

Keywords

reflection-ai, beam-model, open-weight-models, mixture-of-experts, nvidia, coding-agents

Download as PDF

Beam is a new open-weight AI model from Reflection AI, a US startup, built for software work like fixing bugs and running multi-step jobs with tools. It uses a “mixture of experts” design, which keeps it cheap to run: the model has 501 billion parameters in total, but only about 23 billion switch on for any given word. Reflection trained it on 23.8 trillion tokens. The weights arrive later this month; an early version is on a waitlist.

Reflection says Beam matches the larger open model GLM-5.2 on reasoning while using three to four times less computing, so a team running coding agents pays less per task. The training run generated 100 million attempts across 10,500 Nvidia GB300 chips over four weeks, drawn from about one million curated coding and science tasks. Reflection calls Beam the Western open-weight answer to cheaper Chinese models like DeepSeek and Kimi.

Five days after the launch, the Financial Times reported that Nvidia is in early talks to buy Reflection or deepen its existing $800 million investment, at a reported $25 billion valuation. One option is an acqui-hire, hiring the staff and licensing the technology instead of buying the whole company, which would avoid a long antitrust review. Talks could still fall apart. Nvidia is the leading maker of AI chips, so owning a model lab would tie its hardware to the software that runs on it.

Read More: When a Government Can Switch Off a Frontier Model, the Hedge Is Local AI

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {Beam: {A} {501B} {Open-Weight} {Model,} and {Why} {Nvidia}
    {Reportedly} {Wants} {Its} {Maker}},
  date = {2026-10-13},
  url = {https://toknow.ai/posts/reflection-ai-beam-501b-open-weight-nvidia/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “Beam: A 501B Open-Weight Model, and Why Nvidia Reportedly Wants Its Maker.” https://toknow.ai/posts/reflection-ai-beam-501b-open-weight-nvidia/.