Reflection AI ships Beam, a 501B open-weight model
A sparse mixture-of-experts model with 23B active parameters and a context window extendable to 1M tokens, licensed Apache 2.0, with weights due later this month.
Reflection AI announced Beam on 5 October, its first open-weight frontier model: a sparse mixture-of-experts design with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic work. Context length tops out at 256K tokens, extended to 1M during midtraining. It will ship under an Apache 2.0 licence.
Reflection pretrained Beam on 23.8 trillion tokens across 6,144 Nvidia GB300 NVL72 GPUs over roughly four weeks, then ran a reinforcement-learning stage on 10,500 GB300 GPUs for another four weeks, generating over 100 million rollouts. The company reports AIME 2026 at 97.8% and GPQA Diamond at 90.5%, and says Beam matches GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. Early access is open now at platform.reflection.ai; weights, the technical report, a model card and developer artefacts are due “later this month”.
Why it matters: a 501B-parameter MoE under a permissive licence gives anyone who wants to self-host or fine-tune a frontier-class coding and agentic model another option to put against Qwen, GLM and DeepSeek once the weights land.
The caveat: every number here is Reflection’s own, and the weights aren’t out yet. There’s no independent benchmark to check the claims against until the release later in October.