Reflection releases Beam, a 501B open-weights MoE under Apache 2.0

Reflection published Beam, an instruction-tuned open model. The card gives 501.3B total and 23.4B active parameters across 52 layers with 256 experts, a 1M-token context and text-only input. Weights come in BF16, FP8 and NVFP4. The stated knowledge cutoff is June 30, 2026.

The card's main claims are token-efficient reasoning, software engineering and agentic tool use. It says Beam reaches comparable quality to leading open models with about three times fewer reasoning tokens. The results table covers SWE-bench variants, Terminal Bench, AIME 2026, GPQA Diamond and MCP Atlas, all reported by Reflection. The technical report is marked as coming, so the claims cannot yet be checked against a method.

The licence is Apache 2.0. The card says it was developed predominantly for English.

The card gives no release date; the date used is the Hugging Face repo creation date.

A permissive licence and a bold efficiency claim. The report will say how bold.

Sources

• Model card on Hugging Face
• Hugging Face repo metadata (creation date, licence, size)

Editor's first pass (AI-assisted)

• Real 4: Apache 2.0 weights in BF16, FP8 and NVFP4, ungated; the report is still 'coming'.
• Useful 3.5: Coding, tool use and reasoning, with a claim of using fewer reasoning tokens.
• Notable 4: A new entrant releasing a 500B-class model under a permissive licence.
• Novel 3: Conventional sparse MoE; token efficiency is the stated angle.