Intel / article

[AINews] Reflection Beam - 501B-A23B American Open Model

Source: latent.space
Date: Published Oct 6, 2026

Context

Reflection introduced Beam in its own post. Beam is a text-only model (one that activates only a subset of its parts for each step) with 501 billion total parameters and 23 billion active. Reflection says it will release weights under the Apache 2.0 license, plus a technical report, later in October; for now access is by waitlist.

The performance claims are Reflection's: 80.9 on , a coding benchmark, and scores comparable to GLM 5.2 on advanced reasoning tests using 3 to 4 times less estimated generation compute, which Reflection calls approximate rather than measured cost. Latent Space's AINews roundup for October 3 to 5, 2026 says newer models such as GLM 5.3, Kimi K3 and Qwen 3.8 Max are generally ahead, so Beam's draw is a US-trained open model. It also reports that Artificial Analysis, with early access, expects Beam to be among the most token-efficient open models for its intelligence, an expectation rather than a published result.

Key takeaways · AI-distilled

Reflection's team says Beam's 23.8T tokens came partly from an OCR pipeline over hundreds of millions of PDFs, followed by a stable RL/OPD run on 10K GB300s with more than 100M rollouts across about 1M tasks.
Reflection's claimed results include 80.9 on SWE-bench Verified and 3 to 4x the inference efficiency of GLM 5.2. A tech report and open-source integrations are promised but had not shipped at the time of the digest.
Independent reads are mixed: Elie Bakouch estimates only about 12% BF16 MFU in pretraining and reads the design as 3:1 interleaved global and sliding-window , while observers place Beam near GLM 5.2 and below DeepSeek V4 Flash on some benchmarks.
The same issue reports a Hugging Face capture proxy that turned 10 unmodified agent harnesses into RL environments. Identical weights scored 62% under Mini-SWE-Agent and 33% under Claude Code, from a run with one task family and one seed.

Terms in this piece · Glossary

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.
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.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
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.

Read the source www.latent.space