Reflection unveils 501B Beam; Apache-2.0 weights are promised later this month

Reflection AI has introduced Beam, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, positioned for coding, reasoning and agentic workloads. The company calls Beam its first open-weight model, but the current release is a preview: access is limited while final red-teaming and evaluation continue. Reflection says the public weights, technical report, model card and developer artifacts will follow later in October, with the weights planned under Apache 2.0.

The training disclosures are unusually detailed for a preview. Reflection reports 23.8 trillion pretraining tokens and says its high-compute reinforcement-learning run generated more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks. During that RL phase, the maximum context length was 256K tokens. The company separately says midtraining extended Beam's effective context length to one million tokens. Reflection also describes a 6,144-GB300 pretraining cluster and extensive infrastructure for asynchronous RL and sandboxed agent environments.

Reflection publishes benchmark tables comparing Beam with other current open models, especially on coding, terminal, reasoning and tool-use tasks. Those numbers are publisher results and remain attributed as such. The key evidence distinction is availability: Beam is not yet a public checkpoint that Open Model Weights can inspect shard by shard. Until the promised repository and license files appear, it should remain a tracked preview rather than a verified registry entry.