Reflection ships Beam's 501B weights under Apache 2.0

Beam's weights are out, and the license is the news.

Reflection AI put Beam on Hugging Face on Oct. 16, SiliconANGLE reports, 11 days after it unveiled the model with access limited to a select group of early users. The model card lists an Apache 2.0 license.

That means you can download it, fine-tune it, build it into a product and charge for it.

A 501B model you can actually license

Beam is a mixture of experts model with about 501 billion total parameters and 23 billion active per token, according to its model card, with a 1 million token context. It comes in BF16, FP8 and NVFP4 versions.

Reflection's own scores on the card:

81.6% on SWE-bench Verified.

It also lists 80.1% on Terminal Bench v2.1 and 90.6% on GPQA Diamond. Those are Reflection's numbers, not independent ones.

Reflection's launch post is candid that Beam is not the strongest open model. On its own chart, Moonshot's Kimi K3 scores 88.3 on Terminal Bench to Beam's 80.1. The pitch is efficiency instead. Reflection says Beam matches GLM-5.2 on advanced reasoning tests with three to four times less inference compute, because it trained Beam with a length penalty that "rewards successful solutions while discouraging unnecessary tokens."

Alongside the weights, SiliconANGLE says Reflection released fine-tuning recipes, a drafter model for speculative decoding, ready configurations for popular serving engines and fused kernels for faster inference.

Why the license beats the leaderboard

Yes, another giant open model. This one matters for who is allowed to use it.

Beam's closest rivals on Reflection's own chart, Kimi K3 and GLM-5.2, come from Chinese labs. Plenty of companies have a rule against running those weights. Beam is a U.S.-built alternative under the most familiar permissive license there is, with the tooling to customize it shipped on day one.

Cheap open tokens are already winning on volume. Vercel's latest numbers, also in today's edition, show where that traffic is going.

What it takes to run it

This is not a laptop model. Amazon says Beam runs through SageMaker JumpStart in two U.S. East regions at launch. Its guidance puts the FP8 version on a p5-class instance with H100 or H200 GPUs, the NVFP4 version on a B200 instance, and BF16 on more than one machine.

For most teams, the realistic path is a hosted provider, or the released recipes on rented GPUs. If you already fine-tune open-weight models, Beam belongs on your shortlist.

Watch the report and the independent runs

The model card still lists its technical report as coming. Watch for independent benchmark runs, which will settle whether 81.6% holds outside Reflection's harness.

And watch Nvidia. The FT reported last week that Nvidia was in talks to buy Reflection, as TopFive covered. Weights already released under Apache 2.0 stay that way whoever ends up owning the company.

Pull it this week. Run your own eval before you trust the card.