Beam
Reflection AI unveiled Beam, its first open-weight model, a 501B parameter mixture-of-experts model with 23B active parameters, aimed at coding and agent tasks.
Beam is Reflection's first model release. It is a text-only mixture-of-experts model trained with heavy reinforcement learning, with a 1 million token context window and pretraining on 23.8 trillion tokens. The company says it matches Z.ai's GLM-5.2 on reasoning benchmarks while using 3 to 4 times less inference compute. Weights and a technical report are due later this month.
- Date
- Monday, 5 October 2026
- Lab
- Reflection AI
- Kind
- model
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total parameters | 501B 23B active per token; GLM-5.2 is about 744B total, 40B active | company |
| Context window | 1 million tokens | company |
| Pretraining tokens | 23.8 trillion | company |
| Inference compute vs GLM-5.2 | 3-4x less company claim, not independently verified | company |
Announced Monday 2026-10-05. Weights and the full technical report are promised later in October, so access is set to the nearest listed value, and the weights are not yet public.
Sources
- dev.ua/en/news/nova-shi-model-beam-1791272365
- www.benzinga.com/markets/tech/26/10/62181953/nvidia-reflection-ai-chinese-models-deepseek-
- www.deccanchronicle.com/amp/technology/nvidia-backed-reflection-unveils-open-ai-model-taki
- inkorr.com/en/amerikanskij-startap-reflection-ai-gotue-vidkritu-model-si-so-konkuruvatime-
This record was checked and corrected against its sources on 6 October 2026. How we check