← News

Models & toolsSources: arXiv · 20 Nov 2026

Beam technical report describes 501B-parameter sparse MoE model for coding and agentic workloads

A technical report on arXiv describes Beam, a sparse Mixture-of-Experts model with 501.3B total parameters, 23.4B active, and a 1M-token context window, aimed at coding, reasoning, and agentic tasks.

In plain words

A Mixture-of-Experts model keeps many parameter groups but activates only a few per request, cutting compute. The Beam report says its model activates 23.4B of 501.3B parameters and handles very long inputs. The authors state it matches frontier reasoning quality while using 3-4x less inference compute than comparable models.

What was announced

According to the report, Beam was pretrained on 23.8T curated tokens and gained its 1M-token context window through a two-stage midtraining phase built on verifiable, correctness-anchored data. Post-training used a fully asynchronous reinforcement learning campaign across 10.5K GB300 GPUs over four weeks, generating over 100M rollouts; the authors call it one of the largest RL runs by any open lab and say capabilities kept improving with more RL compute without plateauing.

On deployment, the report credits deep hardware-model co-design, FP8/NVFP4 quantization, and a 1.93B-parameter speculative decoding drafter called DFlash 2 with maximizing serving throughput. On safety behavior, the authors report Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless ones, averaged across conversation turns and harm categories, which they describe as the highest harmful-request refusal rate among open-weight models on the Pareto frontier. These figures come from the authors' own evaluation suite and have not been independently verified.

What it does not show

  • Only the paper abstract is available, so the full report's methods and results cannot be checked.
  • All performance, training-scale, and safety figures are author-reported and lack independent confirmation.

For security and IT teams

Scope
The source covers the model's abstract only; deployment details such as weights availability or licensing are not addressed in the supplied text.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.