Open Source · OCTOBER 18, 2026
Reflection ships Beam 1.0 weights under Apache 2.0, Baseten lists input at $1.30/M
The 501B-parameter MoE lands on Hugging Face in BF16, FP8, and NVFP4 with a speculative-decoding drafter, Mirror CLI, and day-0 hosted pricing — turning Reflection's October 5 claim into weights any developer can pull, quantize, or fine-tune commercially.
Reflection AI pushed the Beam 1.0 weights to Hugging Face on October 16, 2026, under Apache 2.0, closing the 11-day gap between its October 5 announcement and a release developers can actually pull, quantize, or fork for commercial use. Baseten listed day-0 Model API pricing at $1.30 per million input tokens, $4.30 per million cache input tokens, and $0.13 per million output tokens, giving the launch an immediate hosted price point against the frontier closed labs.
Beam is a 501.3-billion-parameter mixture-of-experts with roughly 23.4 billion parameters active per token and a 1,000,000-token context window. Weights ship in three precisions: BF16 at 933.81 GiB, FP8 at 468.72 GiB, and NVFP4 at 273 GiB, with the quantized variants fitting on a single eight-GPU server across Nvidia H100, H200, and GB300 configurations. Day-zero serving support covers Nvidia Dynamo, Transformers, vLLM, SGLang, Unsloth, and Axolotl, alongside the Mirror CLI agent harness Reflection open-sourced with the drop.
The benchmark story is self-reported and slightly inconsistent across surfaces. Reflection's model card lists 81.6% on SWE-bench Verified at BF16, 80.1% on Terminal-Bench 2.1, and 97.7% on AIME 2026; Baseten's launch post cites 80.9% on SWE-bench Verified, 90.5% on GPQA Diamond, and 78.7% on MCP Atlas. VentureBeat's read of Reflection's own chart puts Beam at roughly 25,000 tokens per Terminal-Bench 2.1 attempt against about 80,000 for Z.ai's GLM-5.2, the basis for the company's claim of a 3x token-efficiency advantage over other open models.
That framing is the real pitch. Reflection CTO Ioannis Antonoglou and VP of Product Joe Spisak aren't arguing Beam prints cheaper tokens than DeepSeek. They're arguing it burns fewer of them to finish a task, which is the only cost metric that matters inside an agent loop. At Baseten's rate, 10 million input tokens a month runs about $13 before output, the kind of number that reframes what "production inference" means for a founder-led shop.
The license is the other half of the story. Apache 2.0 permits commercial modification without the revenue cliff attached to Meta's Muse Spark Open license, which requires a separate agreement above $20 million in revenue. For teams already routing between Claude and GPT, Beam is a third lane with no vendor lock-in and no per-seat tax.
The usual caveat applies: every benchmark here's Reflection's. Independent evals from Epoch AI or Artificial Analysis will decide whether the token-efficiency chart holds outside the lab. Until then, Beam is a credible, permissively-licensed workhorse with a plausible cost story, released by a company that understands the only durable moat in open weights is being the one developers actually pull.
For context on how efficiency-first architectures are reshaping agent economics, see our coverage of DeepSeek's causal encoder-decoder cost reductions and Mistral Large 4 Le Chonk's benchmark preview.
Sources
- https://www.unite.ai/reflection-releases-beams-open-weights-under-apache-2-0-license/
- https://venturebeat.com/technology/reflections-beam-aims-to-undercut-chinese-open-models-with-fewer-tokens-not-cheaper-tokens
- https://www.datastudios.org/post/reflection-beam-501b-23b-open-weights-1m-context-coding-benchmarks
- https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/
- https://xenospectrum.com/en/reflection-beam-open-model-efficiency/