Beam: Reflection's 501B open-weight model
- Source
- reflection.ai
- Author
- Philpax
- Date
Reflection has shipped its first model after more than a year in stealth. Beam is a text-only sparse MoE aimed at coding, agentic and scientific work, with 501B total parameters and 23B active per token. Team posts collected by Latent Space cite 23.8T pretraining tokens, partly drawn from an OCR pipeline over hundreds of millions of PDFs, and a reinforcement learning run on about 10,000 GB300s with more than 100M rollouts. Reflection's own claims include 80.9 on SWE-bench Verified and three to four times the inference efficiency of GLM 5.2. Those figures are the lab's, not independent measurements. Latent Space notes that Reflection compares Beam with GLM 5.2 and Nemotron, while GLM 5.3, Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash are generally ahead. The practical point is provenance and license: teams that need US-trained open weights under Apache 2.0 have had few choices at this size. With weights not yet published, the useful step now is to prepare an evaluation on your own tasks and serving stack, then compare against the open models you already run once the release and technical report land.
Context
Reflection AI's announcement of Beam describes its first open-weight model: a 501-billion-parameter system, meaning only part of the network (23 billion parameters) is used for each token. The post says Beam is text-only and extends its effective to one million tokens, and it reports results on coding, agentic and reasoning tasks.
The post says Beam is in final red-teaming, with early access by waitlist, and that weights under Apache 2.0 and a technical report are due later in October 2026. The numbers are the company's own: 80.9 on and 90.5 on GPQA Diamond, the latter below GLM 5.2's 91.2 in its table. On advanced reasoning benchmarks, Reflection says Beam's scores are comparable to Z.ai's GLM-5.2 while using three to four times less inference compute. That figure is an estimate from active parameters and average generated tokens, excluding prompt processing, and serving overhead. The page was updated October 8, 2026 with revised Beam results, and until the weights arrive the claims cannot be independently tested.
- mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
- attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
Checking sign-in…
Loading comments…