Reflection releases 501-billion-parameter open-source Beam model

Summary

US-based AI lab Reflection has released the Beam model. This 501-billion parameter mixture-of-experts (MoE) model activates 23 billion parameters per token. It features a 1-million-token context window, an Apache 2.0 license, and a focus on coding and agentic workloads. Additionally, it offers day-one support on transformers, vLLM, and SGLang platforms.

Why it matters

An open-weights model with 501 billion parameters under the Apache 2.0 license, focused on frontier-scale coding and agent tasks, lowers the barrier for teams that do not want to rely on closed APIs. These teams can own, fine-tune, and deploy a large model on their own infrastructure. The model’s hybrid sliding window/global attention mechanism design and its new fixed-scale incremental residual stream architecture are architectural choices that may influence how long-context MoE models are built in the future. The $10,000 Hub credit competition and day-one ecosystem support are read as a deliberate move to build an open-model community around a single new lab’s flagship.

Highlights

• 501 billion total / 23 billion active parameter MoE; 256 experts per layer, 8 routing per token, approximately 1 TB BF16 checkpoint
• 1 million token context with 43 sliding window + 9 global attention layers; reduces key-value (KV) cache growth to approximately 36 KiB per token
• Self-reported benchmarks: SWE-Bench Pro v2-Hard 76.5–78.6%, Terminal-Bench v2.1 78.2–80.1%, AIME 2026 97.7–98

Critical take

All benchmark figures were self-reported by Reflection. There is no independent verification, direct comparison with other open models of similar active parameter count, or ablation study isolating the contribution of the incremental residual stream and query scaling innovations. The text reads more like a launch announcement and marketing tool than a peer-reviewed technical report. The $10,000 competition and demo prompts reinforce this impression. Consequently, the architectural claims have not yet been validated by the community.