October 16, 2026 — Reflection has published the downloadable weights for Beam, its first open-weight model, on Hugging Face under the permissive Apache 2.0 license, turning the October 5 preview into a model anyone can run, inspect and fine-tune. Why it matters: the U.S. open-weight race has mostly been about announcements; Beam is now an actual artifact that banks, governments and builders who will not use Chinese models can download today.
What was released
Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active per token, with a context window of up to one million tokens. Reflection says it was pretrained on 23.8 trillion tokens and then refined with a large reinforcement learning run. The download ships in BF16, FP8 and NVFP4 formats alongside a smaller speculative-decoding drafter model, plus a model card with final benchmark results and a separate safety technical report describing the company’s evaluation methods.
Reflection also opened a public beta of its own API, described as free to start, and open-sourced Mirror CLI, a terminal agent harness it says was built alongside the model for coding and tool-using work.
The numbers — company-reported
The benchmark figures released with the weights come from Reflection’s own evaluations, relayed by launch partner Baseten, and have not been independently verified. On those numbers, Beam scores 90.5% on GPQA Diamond, 80.9% on SWE-bench Verified, 97.8% on AIME 2026, 80.1% on Terminal-Bench 2.1 and 78.7% on MCP Atlas. That is consistent with the positioning from Beam’s October 5 debut: roughly in the class of China’s GLM-5.2 on several coding and agent tests, with Reflection’s pitch resting on efficiency — doing comparable work with fewer active parameters and fewer tokens — rather than topping every leaderboard.
Where you can run it
Availability is broad on day one. Baseten offers Beam through its Model APIs and Dedicated Inference, and Reflection lists launch support across AWS, Microsoft Azure, Google Cloud, Cloudflare, CoreWeave, Databricks, Fireworks AI, LM Studio, Nebius, NVIDIA NIM, Ollama, OpenRouter, Oracle, Together AI, Vercel and Weights & Biases, among others — more than 40 partners and integrations in total. Dell and HPE handle on-premises, sovereign and air-gapped deployments. On the tooling side, the model has day-zero support in Hugging Face Transformers, vLLM, SGLang, Unsloth and Axolotl, so existing self-hosting stacks can pick it up without custom work.
Why it matters
An Apache 2.0 license is the real headline. Unlike custom open-weight licenses with revenue thresholds or usage carve-outs, Apache 2.0 lets a company download Beam, modify it and ship products on top without negotiating with Reflection first. That is exactly the offer Reflection has been making to regulated buyers — an American open model they can run behind their own firewall, on their own data. Whether Beam earns that spot now depends on independent testing, but the barrier to trying it just dropped to a download. Everything AI, tested: expect third-party benchmark runs within days, and those, not the launch table, will decide Beam’s reputation.
Frequently asked questions
Where can I download Beam?
The weights are on Hugging Face under the Apache 2.0 license, in BF16, FP8 and NVFP4 formats. Managed access is available through Baseten and other launch platforms, plus a public beta of Reflection’s own API.
How big is Beam?
It has 501 billion total parameters with 23 billion active per token, a context window of up to one million tokens, and text-only input and output.
Are Beam’s benchmark scores independently verified?
Not yet. The figures published at release are from Reflection’s own evaluations. Independent results should follow now that the weights are public.
What else did Reflection release with the model?
A model card, a safety technical report, a speculative-decoding drafter model, and Mirror CLI, an open-source terminal agent harness.
Sources: Unite.AI, Reflection availability announcement via Unite.AI, Baseten via Unite.AI, Dell