Reflection previews Beam, a 501B-parameter open-weight coding model
Reflection previews Beam, a 501B-parameter open-weight coding model
2026-10-05Verified announcementTRIAL
Beam is now available as Apache-2.0 BF16, FP8 and NVFP4 weights, with Transformers, vLLM and SGLang paths plus a hosted Hugging Face endpoint recipe. The sparse model has 501B total and 23B active parameters with 1M context. Reference hosted deployment uses eight H200 GPUs at about $40 per hour; benchmark claims remain vendor-run.
Availability: Apache-2.0 BF16, FP8 and NVFP4 weights; Transformers, vLLM, SGLang and hosted endpoint
What this means for you
Open weights make Beam independently testable, but its hardware cost is material and vendor benchmarks still need reproduction.
Evidence & related material
Beam moves from limited preview to Apache-2.0 open weights
Subsequent changes
Beam moves from limited preview to Apache-2.0 open weights
Technical details & scores
| Architecture | sparse MoE; 501B total and 23B active parameters |
|---|---|
| Context | 1M effective tokens |
| Training | 23.8T pretraining tokens; over 100M RL rollouts on 10.5K NVIDIA GB300 GPUs over four weeks |
| Availability | Apache-2.0 BF16, FP8 and NVFP4 weights; Transformers, vLLM, SGLang and hosted endpoint |
| License Plan | Apache 2.0 |
Technical impact 4
Engineering value 3
Adoption signal 2
Maturity 3
Verification cost 4
Risk 3