Reflection introduced Beam, its first open-weight model. Beam is a 501B-parameter sparse Mixture-of-Experts model with 23B active parameters. It targets coding, reasoning, and agentic workloads. Pretraining used 23.8T high-quality tokens from web and licensed datasets. Reinforcement learning ran on 10.5K NVIDIA GB300 GPUs for four weeks. It generated over 100M rollouts with 256K-token context. Beam matches or outperforms similar open models on coding and agentic tasks. It achieves comparable scores to larger models while using 3–4× less inference compute. Early access is available via signup. Weights, technical report, model card, and developer artifacts will release later this month.
What changed
No prior Reflection open-weight model existed.
What it unlocks
Running coding and agentic workloads with 3–4× less inference compute than similar models.
- 501B total parameters, 23B active
- 23.8T pretraining tokens
- 100M RL rollouts
- 10.5K NVIDIA GB300 GPUs for 4 weeks
What you need to act on it
- early access signup
- final red-teaming and evaluations pending
Sources