Reflection AI released Beam, a 501B-parameter Apache 2.0 open-weight model for coding and agents

Archive entry: this covers an event from October 16, 2026, and we published it on November 8, 2026. The details come from sources dated at the time and may have changed since.

On October 16, 2026, Reflection AI released the weights of Beam, its first open-weight model, on Hugging Face under Apache 2.0, as announced in Reflection's Beam availability post. Beam is a mixture-of-experts model built for coding, reasoning and long-horizon agents, and the company first announced it on October 5.

What changed

• Beam had 501B total and 23B active parameters. The Hugging Face blog described it as text-only with a 1M context, 52 layers and 256 experts, 8 of them routed.
• BF16, FP8 and NVFP4 checkpoints and a speculative drafter were provided, and the BF16 weights were about 1 TB.
• Reasoning levels were low, medium (the default), high, xhigh and max.
• Day-0 support covered Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo.
• Reflection also open-sourced Mirror, a CLI harness with a Python frontend and Rust core that offers resumable conversations, parallel agents in worktrees, and skills, MCP and Python extensions.
• The Reflection API opened in public beta at platform.reflection.ai and was free to start.

What it meant for people building with AI

Developers got a large Apache-licensed coding model with day-0 support in common inference stacks and its own open agent harness. Reflection said Beam was tested with Pi, OpenCode and Hermes. Scores reported for Beam included 80.9 on SWEBench Verified, 65.5 on SWE Bench Pro v1, 80.1 on Terminal Bench v2.1 and 90.5 on GPQA Diamond.

Platforms, training and sources

Launch platforms included AWS, Baseten, Cloudflare, CoreWeave, Databricks, Fireworks AI, Google Cloud, LM Studio, Microsoft Foundry, NVIDIA NIMs, Ollama, OpenRouter, Together AI and Vercel, with Dell and HPE for on-prem. Reflection's October 5 announcement said Beam was pretrained on 23.8T tokens, then trained with RL on 100M+ rollouts using 10.5K GB300 GPUs over 4 weeks. More detail is on the Beam model page on Hugging Face and in the Hugging Face blog post on Beam.

Sources

• Beam is now available - Reflection (Reflection AI, October 16, 2026, official)
• Introducing Beam: Reflection’s 501B open-weight model - Reflection (Reflection AI, October 5, 2026, official)
• reflection/Beam-1.0-501B-A23B (Reflection AI, official)
• Welcome Beam: a 501B open model you can own, customize, and scale (Hugging Face, October 16, 2026)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.