Reflection AI launches Beam, a 501B/23B-active open-weight model, positioning itself as a new US frontier-adjacent open lab
2026-10-05
Reflection AI announced Beam, a text-only 501B-total/23B-active-parameter mixture-of-experts model for coding, agentic, and scientific work, trained from scratch on a reported 23.8T pretraining tokens plus reinforcement learning on roughly 10,000 GB300 GPUs. Full weights under an Apache 2.0 license are promised later this month. Reflection reports 80.9 on SWE-bench Verified and claims 3-4x the inference efficiency of GLM 5.2. Axios reported ahead of launch that Reflection pays $150M/month for Colossus compute plus a $1B Nebius deal, and that other unnamed US labs plan to ship open models this month. Per Latent Space's roundup of independent observers, Beam lands around GLM-5.2 level, generally behind SOTA open models GLM 5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash.
Significance 4: A new, well-capitalized US lab entering the open-weight frontier race is a significant new entrant per the rubric, though its benchmark claims are lab-reported (capping significance at 4) and independent observers already place it behind current open-weight SOTA.
Implications · machine-drafted, not owner judgment
A credible US-trained-from-scratch open-weight entrant, pitched at enterprises and sovereign nations wanting to fine-tune on proprietary data ('AI factories'), somewhat counters the narrative that open frontier models are exclusively a Chinese-lab phenomenon — worth tracking for the open-weights gap and for sovereign-AI consulting conversations, since Reflection's enterprise-customization pitch overlaps with org-readiness advisory territory.
- Independent SWE-bench Verified reproduction of the 80.9 claim
- Actual Apache 2.0 weight release this month as promised
- Whether other unnamed US labs also ship open models in October
Sources
- press Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost retrieved 2026-10-06
- curator [AINews] Reflection Beam - 501B-A23B American Open Model retrieved 2026-10-06
Affected axes
Coding · Agentic / tool use