Reflection Beam weights are live: 501B/23B MoE under Apache 2.0, BF16/FP8/NVFP4 on Hugging Face, 80.1% Terminal-Bench 2.1

Background

When Reflection announced Beam on Oct 5, weights were still in red-teaming with waitlist-only access. Teams that need a Western open coding model on their own iron were waiting for downloadable checkpoints, not another API tease. On 2026-10-16 the Hugging Face Beam launch post confirmed Apache 2.0 weights are live.

Architecture

Beam is a text-only sparse MoE: ~501B total / ~23B active per token; 52 layers (3 dense + 49 MoE with 256 experts, top-8); 43 sliding-window + 9 global attention layers; 1M context. Depth-scaled residuals keep stream norms stable. Checkpoints: BF16 (~1TB), FP8, and Blackwell-oriented NVFP4 (~280GB).

Benchmarks

The model card (BF16, temp=0.7) reports SWE-bench Verified 81.6%, Terminal-Bench 2.1 80.1%, DeepSWE v1.1 43.3%, MCP Atlas 80.8%, GPQA Diamond 90.6%, AIME 2026 97.7%—vendor figures; re-measure in your stack. Versus multimodal open peers like Muse Spark Open, Beam stays text-only and stresses token efficiency.

Getting started

Primary repo: reflection/Beam-1.0-501B-A23B. Need transformers>=5.20.1; vLLM/SGLang ship day-0 recipes with optional DFlash2. Hardware: plan ≥4×B200 for NVFP4; use FP8 on Hopper without Blackwell. Or try HF Inference Providers (Baseten, Fireworks, Together). HF is offering up to $10,000 Hub credits for strong Beam projects shared on the Hub.

Same day, Reflection open-sourced Mirror CLI (Apache 2.0, beta), its terminal coding agent: uv tool install reflection-mirror, then run mirror in a project. It defaults to Beam but can switch to other providers or custom endpoints; /worktree gives parallel tasks separate checkouts, and sessions persist in ~/.mirror. Tools run with your account's permissions, and manual approval is not a sandbox (docs).