Beam
Reflection AI's US-built open-weight MoE, pitched as an answer to DeepSeek, Qwen and GLM
Context window
1M
Input / 1M tokens
Free
Output / 1M tokens
Free
Provider
Reflection AI
Weights, technical report and model card are promised later in October 2026 under Apache 2.0; early access by waitlist. API pricing not disclosed. · Data verified 2026-10-08
Beam (announced October 5, 2026) is a 501B-total, 23B-active sparse Mixture-of-Experts text model from Reflection AI, pretrained on 23.8T tokens with a 1M-token effective context. Reflection claims parity with GLM-5.2 at 3-4x less inference compute. Its benchmark claims have not yet been independently verified; Artificial Analysis says it is benchmarking the model.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Early access by waitlist; open weights expected later in October 2026.
Strengths
- ✓US-built open-weight frontier contender (Apache 2.0 promised)
- ✓Claims GLM-5.2-level quality at 3-4x lower inference compute
- ✓1M-token effective context
Best for developers who...
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| SWE-Bench Pro v2-Hard | 77.2 | Reflection-reported, not independently verified |
| Terminal-Bench v2.1 | 80.1 | Reflection-reported |
| GPQA Diamond | 90.5 | Reflection-reported |
Source: Reflection AI announcement
Compare Beam
Compare Beam with any other model
Build a comparison →All model comparisons →