Reflection AI releases Beam weights: a 501B open model under Apache 2.0 with 1M context

What Reflection AI released

Reflection AI published the weights of Beam 1.0 on Hugging Face on 16 October 2026 under Apache 2.0. The release post lists BF16, FP8 and NVFP4 weights plus a speculative decoding drafter, a model card with final benchmarks, a Beam Safety Technical Report, a public beta of the Reflection API, developer documentation and Mirror CLI, an open-source terminal harness for running Beam. Reflection says its platform is free to start, but it gives no paid rates.

The model card lists 501.3B total and 23.4B active parameters. The network has 52 layers, 3 dense and 49 mixture-of-experts, with 256 experts and gated grouped-query attention mixed with sliding-window attention. The context window is 1M tokens and the vocabulary has 152k tokens. Beam takes text input and produces text output, was developed predominantly for English, and has a knowledge cutoff of 30 June 2026.

How Beam was trained and how it scores

In its 5 October announcement, Reflection says Beam was pretrained on 23.8T tokens using 6,144 NVIDIA GB300 NVL72 GPUs in under four weeks. Reinforcement learning ran on about 10,500 GB300 GPUs for four weeks with more than 100M rollouts. RL used a maximum context of 256K tokens, and mid-training extends the effective context to 1M.

On the model card, the BF16 weights score 81.6% on SWE-bench Verified, 66.1% on SWE-Bench Pro v1, 80.1% on Terminal Bench v2.1, 97.7% on AIME 2026, 90.6% on GPQA Diamond and 51.1% on Humanity's Last Exam with tools. The FP8 and NVFP4 versions score within about two points. In the announcement table, Beam trails GLM 5.2 at 81.0 and DeepSeek V4.1 Flash at 90.6 on Terminal Bench v2.1, and Reflection says Kimi K3 remains ahead on raw capability. Reflection claims Beam needs three to four times less inference compute than GLM-5.2 on advanced reasoning benchmarks. These are vendor results.

What changed for developers

Until 16 October, Beam was available only through a waitlist, and Apache 2.0 was a stated plan. The weights can now be downloaded, run, modified and redistributed commercially under Apache 2.0, without the in-product notice or the revenue threshold required by Meta's Muse Model License. Reflection lists day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo, and availability on AWS, Google Cloud, Microsoft Foundry, Oracle, CoreWeave, Databricks, Together AI, OpenRouter, Ollama and LM Studio, with on-premises options from Dell and HPE.

What to watch next

Reflection says it will publish self-distillation recipes in the coming days. Per-token API prices and independent benchmark runs are the open questions, since the scores above come from Reflection. The release is listed on the AI model release tracker, and other open-weight launches are collected under Open Source AI.

Questions people ask

Is Reflection AI's Beam open source?

Its weights are released under Apache 2.0, which allows commercial use, modification and redistribution. Reflection published BF16, FP8 and NVFP4 versions on Hugging Face on 16 October 2026.

How big is Beam?

501.3B total parameters with 23.4B active per token, a 1M-token context window and text-only input and output, according to the model card.

How much does the Reflection API cost?

Reflection says its platform is free to start and in public beta. Prices per 1M tokens are not stated in the announcement.

Sources

1. Beam is now available (Reflection AI, 16 Oct 2026)
2. Beam-1.0-501B-A23B model card (Reflection AI on Hugging Face, 16 Oct 2026)
3. Introducing Beam: Reflection's 501B open-weight model (Reflection AI, 5 Oct 2026)