Reflection’s Beam arrives on Hugging Face as a 501B Apache 2.0 open model

Hugging Face introduced Beam, the first model from new U.S. lab Reflection: a 501-billion-parameter mixture-of-experts with about 23 billion parameters active per token, text-only, with a 1-million-token context and an Apache 2.0 license. Checkpoints come in BF16, FP8 and NVFP4. Day-zero support spans transformers, vLLM and SGLang, and a Rust-based harness ships with the model. Reflection is offering $10,000 in Hub credits for strong community projects, and Inference Providers and Endpoints recipes are live. Reflection earlier framed Beam as the most efficient model in its class rather than the strongest (Reflection introduces Beam).

Reflection-provided benchmarks across precisions include 81.6% on SWE-bench Verified in BF16, 80.1% on Terminal Bench v2.1, 80.8% on MCP Atlas and 97.7% on AIME 2026. Only 8 of 256 experts run per token, but all experts stay in memory, so deployment memory tracks the full 501B while compute tracks about 23B. Hardware guidance favors NVFP4 on Blackwell, with plans for multiple B200 GPUs.

Apache 2.0 on a 501B coding and agent model is a significant open release: users can own, customize and serve it without restrictive terms. The post’s notes on residual-stream and hybrid-attention design give practitioners useful detail beyond a weight dump. Beam arrives alongside Meta’s 816B Muse Spark Open (Muse Spark Open) and StepFun’s 600B Step 5 Preview (Step 5 Preview), showing open labs competing on agentic coding at frontier scale. For enterprises, it offers a path to private coding agents via self-hosted FP8 or NVFP4 or Hugging Face endpoints.

The leaderboard figures are vendor numbers until outsiders reproduce them. Beam is useful for teams with the GPUs to run it; hosted inference providers lower that barrier for experimentation.

Independent SWE-bench and Terminal Bench runs will show whether Reflection’s table holds, and whether the Rust harness proves stable in production agent loops. Community projects chasing the Hub credits will indicate developer interest. For closed API vendors, more Apache-licensed competition on agentic coding adds price pressure, as decision-model and open-weight adoption on gateways already suggests (Vercel gateway data). More in our AI coverage, tools coverage and startups coverage.