ReflectionAI releases Beam, an open-source model pitched as alternative to Chinese LLMs



Nvidia-backed startup ReflectionAI has introduced Beam, a 501B-parameter open-source model with 23B active parameters that reportedly rivals GLM-5.2 while being 3–4x more efficient at inference.
ReflectionAI, a startup funded by Nvidia, has released Beam, a powerful open-source AI model presented as an alternative to Chinese open-weight LLMs.
By metrics, Beam appears to be roughly on par with GLM-5.2. The team claims it is 3–4 times more efficient during inference, saving tokens and completing tasks faster and cheaper.
Technical details:
- 501B parameters, 23B active parameters.
- The model was trained from scratch; pretraining took one month on 24T tokens.
- ReflectionAI ran the largest previously documented reinforcement-learning (RL) run: four weeks on 10.5k GB300 chips. The company reportedly rents the chips for $150M per month.
Weights are promised “soon” under the Apache 2.0 license: reflection.ai/blog/introducing-beam.
A full replacement for Chinese open-source models has not happened yet, but this is the startup’s first model, so the results are already notable.
Why it matters
AnalysisBeam shows that Western startups can build large open-weight models that compete with China’s open-source ecosystem and lower inference costs. If Apache 2.0 weights arrive, developers will gain another option for local and cloud deployment.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
AI Digest: NVIDIA CEO Rumors, Gemini, Kimi, MiniMax, and OpenAI Updates
Key AI developments from November 2–15: rumors of a Jensen Huang leadership shake-up, new Gemini and Kimi models, MiniMax image generation, and OpenAI's Codex retro console.
OpenMotion: An Open-Source CapCut Alternative for AI-Agent Video Editing
OpenMotion has appeared on GitHub as a full-featured CapCut alternative that creates and edits videos using AI agents, templates, and brand kits.
Open Weights vs Truly Open Source AI
A legal and engineering analysis of the fundamental difference between accessible numerical weights of models (Llama, DeepSeek) and fully open projects with source data, code, and architecture (OSI Standard).
Jev and System One AI: A Complete Practical Guide to the Decision Model from TypeSafe
A deep engineering breakdown of the flagship Jev model: System One architecture versus LLMs, Choice/Score/Noul primitives, request parallelism, confidence mathematics, video use cases, code, and an overview of local open-weight alternatives (Laya, GLiNER2.5, CLM-8B).