Reflection AI releases Beam weights under Apache 2.0: 501B-A23B MoE with 1M context in BF16/FP8/NVFP4, model card puts SWE-bench Verified at 81.6%, 40+ partners at launch

Why it matters

【Key points】

On Oct 16 (13:01 UTC) Reflection AI made Beam generally available: a 501B-total, 23B-active sparse MoE whose weights are now on Hugging Face under Apache 2.0, in BF16, FP8 and NVFP4 plus a DFlash2 speculative-decoding drafter, alongside a model card and a Beam Safety Technical Report. This delivers the weights promised when Beam was previewed on Oct 5.

Model card: 52 layers (3 dense + 49 MoE, 256 experts), 1M context, text only, knowledge cutoff June 30, 2026. Self-reported BF16 scores: SWE-bench Verified 81.6%, Terminal Bench v2.1 80.1%, SWE Bench Pro v2-Hard 76.5%, GPQA Diamond 90.6%, AIME 2026 97.7%, HLE (no tools) 36.0%, MCP Atlas 80.8%, BrowseComp 76.8%. FP8 and NVFP4 mostly land within 1–2 points.

Around it: the open-source Mirror CLI coding agent; day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl and NVIDIA Dynamo; 40+ partners including AWS, Google Cloud, Azure, OpenRouter and Ollama; Dell and HPE cover on-prem and air-gapped installs. Reflection's in-house Iris safety suite: 88.9% refuse-or-pivot on harmful requests, 81.5% compliance on harmless ones.

【Verifiable】

Blog: https://reflection.ai/blog/beam-is-available

Model card: https://reflection.ai/beam-model-card

Weights: https://huggingface.co/reflection/Beam-1.0-501B-A23B (public, ungated, apache-2.0)

Announcement: https://x.com/reflection_ai/status/2111079965529968798

【Our read】

1. A second major US open-weight drop this week after Meta's Muse Spark Open. Reflection turned a preview into downloadable, commercially usable Apache 2.0 weights, a looser licence than most Chinese labs' custom ones.

2. Final scores track the preview, some up, some down: SWE-bench Verified 80.9 to 81.6, SWE Bench Pro v2-Hard 77.2 to 76.5, DeepSWE 44.4 to 43.3. All self-reported; await independent runs.

3. The real bet is the bundle of weights, tooling and 40-partner distribution. Watch whether the promised open-sourcing of the Iris safety suite and the self-distillation recipes ship on time.