Reflection releases Beam's weights under Apache 2.0: 501B open model, 23B active, free on its API for now

Reflection released the weights of Beam, its first open-weight model, on Hugging Face on October 16 under the Apache 2.0 license, eleven days after announcing it. Reflection's post "Beam is now available" went up the same day, its account announced it on X at 13:01 UTC, and our launch watcher fired at 13:47 UTC.

WHAT SHIPPED

The instruction-tuned model in three precisions, BF16, FP8 and NVFP4, plus a DFlash2 drafter for speculative decoding, the tokenizer and a reference implementation, all under reflection on Hugging Face. Reflection also open-sourced Mirror CLI, the coding harness it says Beam was trained with.

THE MODEL

A sparse mixture-of-experts with 501.3B total and 23.4B active parameters, 256 experts and 52 layers, text in and text out, a 1M-token context window and up to 128K output tokens, by Reflection's model card and API docs. Knowledge cutoff June 30, 2026. The card adds no use restrictions on top of Apache 2.0: Reflection says it does not prohibit distillation, training on outputs, fine-tuning or merging.

THE PRICE

None yet. Reflection's own API is a free public beta, and Vercel's AI Gateway lists reflection/beam at $0 / $0 per million tokens. Reflection names more than 20 hosts at launch, OpenRouter among them, but OpenRouter's catalog had no Beam id when we checked at 13:58 UTC.

HOW IT SCORES

Artificial Analysis already has a page for it: 30 on its Intelligence Index on launch day. On the same index, as our benchmark table read it today, GLM-5.3 is at 45, Kimi K3 at 44, Mistral Large 4 Preview at 38 and DeepSeek V4 Pro at 36, so Beam starts behind the leading open models from China and behind Mistral's.

Reflection's own figures, BF16: SWE-bench Verified 81.6%, Terminal Bench 2.1 80.1%, GPQA Diamond 90.6%, AIME 2026 97.7%, Humanity's Last Exam 36.0% without tools and 51.1% with them. Reflection says Beam matches leading open models with about 3x fewer reasoning tokens. Those are the vendor's numbers, not a finding.

THE STEALTH NAMES

Reflection's post says early developers used Beam under two codenames: "Catjam" through OpenCode and "MissingNo" through Nous Portal. We had cat-jam on our radar since October 13 as a free stealth model with no named owner. It now resolves to Beam.

WHAT WE DO NOT KNOW

What Beam will cost once the free beta ends, and what the other boards make of it. Reflection says a technical report and reproducible eval recipes are coming. The Nvidia talks the Financial Times reported on October 10 have not produced an announced deal.