Reflection AI launches Beam, a 501B-parameter open-weight MoE model for coding and agents

On October 5, AI lab Reflection AI released its new flagship model Beam, a sparse mixture-of-experts (MoE) model with 501B total parameters and 23B active parameters per inference, targeting coding, reasoning and agentic workloads.

Beam was pretrained on 6,144 NVIDIA GB300 NVL72 GPUs in under four weeks, processing 23.8 trillion tokens with a peak goodput of 92.3%. The reinforcement learning stage used an additional 10,500 GB300 GPUs over four weeks to complete over 100 million rollouts, with about 1.3 billion sandboxes and nearly 1 million high-quality coding and agentic environments. The model has a maximum context length of 256K tokens, with midtraining extending the effective context to 1M tokens.

On capability benchmarks, Reflection AI says Beam matches or slightly exceeds GLM 5.2 on coding and agentic tasks while approaching Qwen 3.8-Max, but uses 3-4x less inference compute than comparably capable models. Beam introduces a tunable reasoning effort parameter, letting users trade off response length and performance. The RL training runs fully asynchronously and remains stable even with policy staleness exceeding 107 weight versions.

On infrastructure, Beam uses interleaved local and global attention with auxiliary-loss-free balancing. Weight distribution hits 99.99% packing utilization, with new weights reaching inference clusters in a median of 12 seconds and 75% less cross-rack traffic. MoE expert utilization is nearly perfectly uniform, with the busiest expert loading at just 1.04x average.

For safety and alignment, the team defines three principles: inviolable safety rules, qualities that should always hold, and a default interaction mode, using Multi-Teacher On-Policy Distillation (MOPD) to combine capability and safety teacher signals. Reflection AI plans to release Beam weights, a model card, technical report, and the full run/eval/finetune toolchain under Apache 2.0 later this month.