Reflection Releases Beam's Open Weights: Apache 2.0 License, 501B-Parameter MoE

What happened

On October 16, Reflection published Beam's model weights on Hugging Face under an Apache 2.0 license, accompanied by a model card and a Beam Safety Technical Report. The release fulfills the company's October 5 promise to publish the weights later in the month. Beam is a text-in, text-out autoregressive mixture-of-experts model: 501 billion total parameters, 23 billion active per token, and a one-million-token context window. Reflection says it pretrained the model on 23.8 trillion tokens and ran reinforcement learning for four weeks across 10,500 NVIDIA GB300 GPUs, generating more than 100 million rollouts.

Key facts

1. Model specs: Text-in, text-out autoregressive MoE: 501 billion total parameters, 23 billion active per token, one-million-token context window
2. Training scale: Pretrained on 23.8 trillion tokens; the reinforcement learning phase ran four weeks on 10,500 NVIDIA GB300 GPUs, generating more than 100 million rollouts
3. Vendor benchmarks: GPQA Diamond 90.5%, SWE-bench Verified 80.9%, AIME 2026 mathematics 97.8%, Terminal-Bench 2.1 80.1%, MCP Atlas 78.7%
4. Pricing: Baseten API per million tokens: $1.30 input, $4.30 cached input, $0.13 output
5. Deployment: Day-one support from 40+ platforms including AWS, Azure, Google Cloud, Baseten, Together AI, Fireworks AI, Nebius, Modal, CoreWeave, Databricks, Cloudflare, NVIDIA NIMs, Ollama, LM Studio, OpenRouter, Oracle, Vercel and W&B; Dell and HPE deliver behind-the-firewall deployments; precision formats BF16/FP8/NVFP4 plus a speculative decoding drafter model
6. Mirror CLI: Also open-sourced a terminal agent harness: a Python frontend with a Rust agent core, supporting resumable conversations, parallel agents in separate worktrees, project instructions, MCP connections and Python extensions
7. First-party API: Public beta at platform.reflection.ai, free to start