Reflection AI released the weights for Beam, its first open-weight model, on Hugging Face under an Apache 2.0 license on October 16. Beam is a sparse mixture-of-experts model with 501 billion total parameters, 23 billion of them active per token, built for coding, reasoning and agentic workloads. Reflection AI first described the model on October 5, while Beam was still in final red-teaming and evaluations.
Reflection ships the weights in BF16, FP8 and NVFP4, plus a speculative decoding drafter model, and lists day-zero support in Hugging Face Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA's Dynamo. Beam is also offered on 22 listed platforms, among them AWS, Google Cloud, Microsoft Azure, Cloudflare, Databricks and Ollama, while Dell and HPE supply installs for on-premises and air-gapped sites. Reflection AI's own API is free to start in a public beta.
Beam runs 52 layers, three dense and 49 mixture-of-experts, Reflection states. Each token is routed to eight of its 256 experts, and the model has a 1M-token context window with text-only input and output. Its reinforcement-learning stage ran at contexts of up to 256K tokens. The knowledge cutoff is June 30, 2026. In the final BF16 results, Beam scores 81.6% on SWE-Bench Verified, 80.1% on Terminal Bench v2.1 and 90.6% on GPQA Diamond. Reflection says it reaches quality comparable to leading open models with roughly three times fewer reasoning tokens.
Reflection AI describes Beam as advancing the Western open-weight frontier and compares it with seven other models. On Terminal Bench 2.1, Beam's 80.1% trails five of them and leads Inkling and Nemotron 3 Ultra. The five range from GLM 5.2 at 81.0% to DeepSeek V4.1 Flash at 90.6%, scores Reflection takes from Artificial Analysis and DataCurve. Reflection's case is efficiency: it says Beam's advanced-reasoning scores are comparable to GLM 5.2's while using three to four times less inference compute. Reflection estimates that compute from active parameters times generated tokens, leaving out prompt prefill, context-dependent attention and serving overhead, so it is not a measured serving cost.
On the tooling side, Reflection trained Beam alongside Mirror CLI, a terminal harness that runs parallel agents in separate worktrees and supports MCP connections, with a Python front end over a Rust agent core that drives the model. Beam has also been tested with Pi, OpenCode and Hermes. Reflection plans to publish self-distillation recipes, which train a LoRA adapter on verified traces from a developer's own tasks and then repeat the loop.
