Reflection AI Launches Beam: 501B-Param Open-Weight MoE With 80.9% SWE-bench and 4-Week GB300 RL Run
Reflection AI announced its first open-weight model Monday, with weights scheduled for public release later in October after red-teaming completes. Beam is a sparse Mixture-of-Experts architecture: 501 billion total parameters, 23 billion active per forward pass. The model targets coding, reasoning, and agentic workloads.
Pretraining covered 23.8 trillion tokens from web and proprietary licensed datasets, a corpus Reflection says matches or outperforms available open base models of comparable size. The RL phase was heavier than most: 100 million rollouts on 10,500 NVIDIA GB300 GPUs run continuously over four weeks.
Benchmark Numbers
| Benchmark | Beam | Inkling | GLM 5.2 | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|---|---|---|
| SWE-bench Verified | 80.9% | 77.6% | — | — | — |
| SWE-bench Pro v1 | 65.5% | 54.3% | 62.1% | — | — |
| Terminal-Bench v2.1 | 80.1% | 63.8% | 81.0% | 88.3% | 90.6% |
| DeepSWE v1.1 | 44.4% | — | 44.0% | 68.0% | 74.2% |
On raw capability, Kimi K3 and DeepSeek V4.1 Flash still lead the open-weight field. Reflection’s pitch is efficiency: Beam achieves comparable reasoning scores to GLM-5.2 while using 3-4x less inference compute measured in FLOPs per generation. Against 2T+ parameter models like Qwen 3.8-Max, which require significantly more compute per token, the gap widens further.
Positioning
Reflection was co-founded by DeepMind researchers. The company has been explicit about its framing: it wants Beam to be the “obvious choice” for enterprise developers who would otherwise reach for Chinese frontier open-weights. Its closest competitive target is the GLM-5.x family and, on coding tasks, an approach toward Qwen 3.8-Max territory.
On SWE-bench Verified, Beam’s 80.9% places it above Thinking Machines’ Inkling (77.6%) and at the same level as Claude Sonnet 4.6 (79.6%) and Kimi K2.6 (80.2%) — all models with significantly different inference economics.
Inference Economics
Beam’s active parameter count of 23B means serving cost per token is comparable to a 23B dense model, not a 501B one. The efficiency claim: frontier-class benchmark scores at mid-tier serving cost. Reflection has not announced API pricing yet.
Weights, a technical report, and a model card will be published later in October. The model is completing red-teaming. Early access sign-ups are open at platform.reflection.ai.