# Beam — Reflection's 501B open-weight MoE for coding and agents

> Reflection AI's first open-weight model: a 501B MoE built for coding, reasoning and agent workloads.

- Org: Reflection AI
- Date: 2026-10-05
- Category: model
- Source: https://reflection.ai/blog/introducing-beam
- Canonical page: https://ai-tldr.dev/releases/reflection-beam/

## In short

Reflection AI announced Beam, a 501B-parameter Mixture-of-Experts model with 23B active parameters for coding, reasoning and agent work. Weights arrive later this month under Apache 2.0; early access is open by waitlist.

## What it is

Beam is Reflection AI's first open-weight language model, announced on 5 October 2026. It is a text model aimed at coding, reasoning and agentic tasks, and Reflection plans to publish its weights under Apache 2.0 later this month.

## How it works

Under the hood, the model is a sparse Mixture-of-Experts with 501B total parameters, of which only 23B are active for each token, which keeps inference cheaper than a dense model of the same size. Reflection trained it on 23.8T tokens, extended its context in midtraining, and ran a four-week reinforcement-learning stage with more than 100 million rollouts.

## Why it matters

A large Apache-2.0 model from a US lab gives companies and governments an open alternative to the leading Chinese open-weight models. If Reflection's efficiency claim holds, teams could run GLM-5.2-class coding agents on a fraction of the hardware.

## Who it's for

teams that self-host open models for coding agents

## Try it

https://platform.reflection.ai

## Quick facts

- Maker: Reflection AI
- Parameters: 501B total, 23B active (sparse MoE)
- Training data: 23.8T tokens
- License: Apache 2.0 (weights later this month)
- Availability: Early-access waitlist

## FAQ

### When can I download the Beam weights?

Reflection AI says it will release the Beam weights under the Apache 2.0 license later in October 2026, together with a technical report, a model card and tools for running, evaluating and fine-tuning the model. Until then, Beam is available only through an early-access waitlist on Reflection's platform.

### How does Beam compare to GLM 5.2 and Qwen 3.8-Max?

Reflection AI claims Beam is competitive with larger open models like GLM 5.2 and approaches Qwen 3.8-Max on coding and agentic tasks. SiliconANGLE reports that Beam does some tasks better than GLM-5.2 while using between one third and one fourth of the hardware. These are the maker's own benchmark results, not independent tests.

### How was Beam trained?

Reflection AI pretrained Beam on 23.8 trillion tokens in under four weeks on a cluster of 6,144 NVIDIA GB300 NVL72 GPUs. For the reinforcement-learning stage, Reflection used 10.5K GB300 GPUs for four weeks and generated more than 100 million rollouts, with about 1.3 billion sandboxes used for training and grading.

### Is Beam good at coding benchmarks?

Reflection AI reports that Beam scores 80.9 on SWE-bench Verified, 80.1 on Terminal Bench v2.1, 78.0 on SWE-bench Multilingual and 77.2 on SWE Bench Pro v2-Hard. Reflection also lists 97.8 on AIME 2026 and 90.5 on GPQA Diamond for reasoning.

## Links

- [Announcement](https://reflection.ai/blog/introducing-beam)
- [Early Access](https://platform.reflection.ai)
- [SiliconANGLE Coverage](https://siliconangle.com/2026/10/05/reflection-ai-debuts-open-source-beam-model-with-501b-parameters/)
- [HN Discussion](https://news.ycombinator.com/item?id=49969183)
