Reflection AI releases Beam, a 501B open-weight model claiming parity with Chinese rivals at lower compute cost
Reflection AI, a Brooklyn-based startup founded by former DeepMind researchers, has unveiled Beam, its first open-weight model. The 501-billion-parameter Mixture-of-Experts model activates 23 billion parameters per token and targets coding, reasoning, and agentic workloads. Reflection claims Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using three to four times less…
Key points
- Beam is a 501B MoE model with 23B active parameters, scoring 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1 per Reflection's benchmarks.
- Reflection claims Beam matches GLM-5.2 on reasoning using 3-4x less inference compute; Apache 2.0 weights coming later this month.
- Sources differ on investors: TechCrunch cites Lightspeed and $4.7B raised; Semafor cites Citigroup and $25B valuation.
Reporting differs on funding details. Semafor and TechCrunch cite a $25 billion pre-money valuation, but TechCrunch puts total capital raised at roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, while Semafor lists investors as Nvidia, Sequoia, and Citigroup. Only TechCrunch and Unite.ai report Reflection's compute deals worth over $7 billion with SpaceX and Nebius for Nvidia GB300 chips through 2029. Unite.ai uniquely details Reflection's sovereign AI cloud partnership with Shinsegae Group in South Korea and its certification with the U.S. Department of Energy's Genesis Mission. Marktechpost provides the most granular technical specifications, including architecture details like SandwichNorm and FP32 residual accumulation, and reports a 92.3% training goodput with nine semi-automatic rewinds.
Model pages: Beam → · Kimi K3 → · Nemotron 3 Ultra →
Reflection unveils Beam, an open model it says rivals GLM 5.2 on reasoning while using 3x-4× less compute and approaches Qwen3.8-Max on coding and agentic tasks
semafor.com · 5 October 2026
Loading the full article…
This text was published by semafor.com and written by https://www.semafor.com/author/jd-capelouto, https://www.semafor.com/author/ashley-gold. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
4sources- Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic WorkloadsPress · MarkTechPost ·
- Reflection AI Unveils Beam, a 501B-Parameter Open-Weight ModelPress · Unite.AI ·
- Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute costPress · TechCrunch AI ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Z.ai's GLM 5.3 model becomes available on Amazon Bedrock · 1 src
- Anthropic prompts users to share voice data to enhance AI models · 6 src
- Reka AI releases research preview of Rho-1 omni-model · 1 src
- Claude explains AI context windows and their limits · 1 src
- OpenAI, Google, and Anthropic launch new AI models in late September · 4 src
Comments
via GitHub Discussions