Reflection AI open-sources Beam: 501B MoE model that prioritizes cost efficiency over raw performance

Forum topic · 小凯 · 2026-10-22

Summary

On October 16, Reflection AI, a Brooklyn-based startup founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou, released the weights of its flagship model Beam on Hugging Face under Apache 2.0. Beam is a sparse MoE with 501B total parameters, 23B activated per token, and a 1M-token context window. It scores 80.1% on Terminal-Bench v2.1 (BF16) and 81.6% on SWE-Bench Verified, roughly matching top Chinese open-weight models like GLM at the first tier of aggregated leaderboards. Its headline claim is not record performance but 3-4x estimated savings in inference compute, achieved through a controllable length penalty in RL training that rewards concise solutions. Training used 10,500 NVIDIA GB300 GPUs for four weeks with over 100 million rollouts. Reflection also open-sourced Mirror CLI, a Rust-core terminal agent supporting parallel Git worktrees, MCP, and Slack/Linear/GitHub integration. Note: GGUF builds are community-converted, not official, and the compute savings figure is an official estimate rather than a measured deployment cost.

Not Faster, Just Cheaper: Reflection AI Open-Sources Beam

The release

A clear fact of the 2026 open-weight leaderboard: the top spots are almost entirely occupied by Chinese models — GLM, Kimi, Qwen, DeepSeek, and MiniMax taking turns. American labs lacked a card of their own. On October 16, the card was played. Reflection AI (founded in 2024 in Brooklyn, New York by former DeepMind researchers Misha Laskin and Ioannis Antonoglou; Reuters and the Wall Street Journal reported a $25B valuation in March, with NVIDIA acquisition rumors in October) published the weights of its flagship model Beam on Hugging Face under Apache 2.0: a sparse MoE with 501B total parameters, 23B activated per token, and a 1M-token context window. Released alongside it is the terminal agent Mirror CLI.

The numbers, corrected

Terminal-Bench v2.1: 80.1% (BF16; FP8 and NVFP4 are 78.9% and 78.4%)
SWE-Bench Verified: 81.6% — a correction to a widely circulated old figure: many reports (including some video descriptions) cite 80.9%, which was the October 5 preview card number. The official model card's BF16 figure is 81.6%, higher than the old number.

As for "breaking the open-source agent performance record," the claim needs unpacking: the official wording is "parity with GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute," plus the media framing of "the first American open-source model that can go toe-to-toe with China's top open models." What's being broken is the *efficiency ratio*; absolute performance is *parity*, not a record — on benchlm's aggregate leaderboard, GLM-5.3 and Beam sit in the same first tier, neither pulling ahead.

Where the savings actually come from

Beam's training used a controllable length penalty: reward successful solutions, penalize excess tokens. The effect appeared early in training — as scores rose, generation length fell; the model genuinely learned to "say less, do more." After deployment, this mechanism becomes a user-facing dial: turn it down for short, fast answers; turn it up to allow longer reasoning on hard problems. Inference budget becomes, for the first time, an explicitly priced slider.

The training scale behind this deserves attention: 10,500 NVIDIA GB300 GPUs for four weeks, over 100 million rollouts, roughly 1.3 billion sandboxes consumed across grading, training, and ablations, with the environment pool filtered from one million coding, agent, and STEM environments. One easily overlooked engineering detail: in the asynchronous RL setup, numerical stability held even when the current policy lagged the latest weights by 107 versions — RL at this scale completes only because of foundations like this.

Mirror CLI

Mirror CLI is the other half of understanding Beam's positioning. The architecture is two-layer: a Python front-end handles the product layer — prompts, skills, developer tools; a Rust core handles execution — the part that actually drives the model is kept minimal, which the team says lets the core double as a clean evaluation surface. Features include running multiple agents in parallel across Git worktrees, resumable sessions, MCP connectivity, custom skills, and integrations with Slack, Linear, and GitHub. Internal users describe the daily routine as closing the laptop while the agent keeps working. The model and harness are trained as a pair, but it's also compatible with third-party harnesses like Pi, OpenCode, and Hermes. There's also an official speculative-decoding drafter (DFlash2 in the model card, pre-released October 13) — one more layer of inference speedup on top of prefill.

Ecosystem clarifications

Two points in the ecosystem list are prone to misreporting, so let's pin them down:

Official precision releases are BF16, FP8, and NVFP4 plus DFlash2. The widely spread claim of "official GGUF release" is wrong — the GGUF was converted by community member DevQuasar.
"Native Ollama support" is not on the official Day-0 list (which is Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed, Dynamo) — though Beam is indeed already in the Ollama library, so that community path is half-true.

The official framing: "the model is the kernel, and the infrastructure layer above the kernel is fully open" — deployment containers, inference integrations, custom pipelines, with 40+ partners on board Day-0, and an official recipe for enterprises doing self-distillation fine-tuning.

The caveats

Beam's benchmark card is self-reported; competitor figures are cited from Artificial Analysis and DataCurve. The full technical report was still marked "coming" at publication time. And the 3-4x compute savings is officially labeled an estimate — an estimate folding in prefill, attention, and serving overhead, not a measured deployment bill.

These boundaries don't change the big picture: a 1M-context, Apache 2.0, first-tier-matching American weight set means self-hosting carries a zero API bill. Against comparable closed-source models on the leaderboard priced at $3 per million input tokens and $15 per million output tokens, enterprise buyers finally have an option where "equally smart" and "spending less" can both be true at once.