Welcome Beam: a 501B open model you can own, customize, and scale
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
Beam is a 501B-parameter Apache 2.0 MoE coding/agent model with only 23B active parameters per token and a 1M-token context window. For production teams, the important shift is that frontier-scale long-context agent workloads can now be self-hosted, customized, and post-trained with day-0 support in transformers, vLLM, SGLang, plus dedicated inference/fine-tuning tooling, instead of being locked behind closed hosted APIs.
Reflection dropped Beam, a 501B-parameter MoE (23B active, 1M context) under Apache 2.0 with day-0 support in transformers, vLLM, and SGLang, plus BF16/FP8/NVFP4 checkpoints and a Rust inference harness — meaning you can self-host a frontier-scale coding/agentic model with no license restrictions and full fine-tuning (SFT/LoRA/RL) rights. The sparse activation keeps serving costs near a 23B model while giving you ownership and customization that closed APIs can't, but you'll need serious multi-GPU hardware to hold 501B weights even at FP8/NVFP4. Tool calling uses a dedicated grammar and reasoning is tunable across five effort levels with cross-turn thinking preservation, so plan your agent harness around those knobs.