Reflection has launched Beam, a 501 billion parameter open-weight model designed for coding and agentic workloads. Beam is currently undergoing final evaluations and will be available for early access soon.
Key facts
| Fact | Detail | The source says |
|---|---|---|
| Total Parameters | 501 billion | “Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active” |
| Active Parameters | 23 billion | “Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active” |
| Pretraining Tokens | 23.8 trillion | “We pretrained the model on 23.8 trillion diverse, curated, high-quality tokens from the web and proprietary licensed datasets” |
What happened
Reflection has introduced Beam, a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion of which are active. Beam is designed for coding, reasoning, and agentic workloads. The model was pretrained on 23.8 trillion tokens from diverse, curated, and high-quality sources, matching or outperforming similar-sized open base models. Reflection also developed the infrastructure needed to sustain high-compute reinforcement learning at exceptional scale, generating over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks of training. Beam is undergoing final red-teaming and evaluations and will be available for early access soon.
What to weigh
- Beam is undergoing final red-teaming and evaluations.
- Beam is designed for coding and agentic workloads.
Source: reflection.ai
