Welcome Beam: a 501B open model you can own, customize, and scale

Hugging Face introduces Beam by Reflection, the first model from the new U.S. lab: a 501-billion-parameter mixture-of-experts with about 23 billion active per token, text-only, 1M context, Apache 2.0. Checkpoints include BF16, FP8, and NVFP4. Day-0 support spans transformers, vLLM, and SGLang; a Rust-based harness ships with the model. Reflection is offering $10,000 in Hub credits for strong community projects.

Benchmarks in the post are Reflection-provided across precisions — e.g., SWE Bench Verified 81.6% BF16, Terminal Bench v2.1 80.1%, MCP Atlas 80.8%, AIME 2026 97.7%. Hardware guidance favors NVFP4 on Blackwell (plan for multiple B200s). Inference Providers and Endpoints recipes are live.

Apache 2.0 on a 501B coding/agent MoE is a serious open release — own it, customize it, serve it. The residual-stream and hybrid-attention design notes are the kind of detail practitioners need, not just a weight dump.

We’ll treat the leaderboard as vendor numbers until outsiders reproduce them. Still, Beam joining Muse Spark Open in the same news cycle means open U.S. labs are competing on agentic coding at frontier scale. Useful if you have the GPUs; irrelevant if you don’t — Inference Providers blunt that for experimentation.

I’m watching independent SWE and Terminal Bench runs, and whether the Rust harness becomes a reason to pick Beam over rivals.

Only 8 of 256 experts run per token; all experts stay in memory, so deploy memory tracks the full 501B while compute tracks ~23B.