The license is Apache 2.0. The weights are on Hugging Face and they are yours. Except even the lightest version of Beam, NVFP4, weighs about 280 gigabytes, and BF16 about a terabyte.
Fact
The facts: on October 16, 2026, Beam, the first model from the new US lab Reflection (per Hugging Face), came out on Hugging Face; Reflection itself lists it as its first open-weight model. Beam is a mixture-of-experts model with 501B parameters, 23B of them active per token: 52 layers, 256 experts per MoE layer, 8 routed per token. It takes text in and gives text out, with a context window of 1M tokens, under the Apache 2.0 license. The weights come in three formats: BF16, FP8 and NVFP4. Per Hugging Face, BF16 is about 1 TB and NVFP4 shrinks it to about 280 GB. Hugging Face recommends four B200s with 720 GB in total and describes Beam as focused on coding and agentic use cases. The benchmark results are from the Reflection team (per Reflection): 81.6% on SWEBench Verified and 90.6% on GPQA Diamond in BF16; the table compares the three formats with each other, not Beam with other models.
Two numbers in the announcement read together. The active parameters are 23 billion, but all 501 sit in memory. Hugging Face says it plainly: memory is that of a 501B model, speed is that of a 23B one.
The weights are really open. The hardware is another matter. The NVFP4 version runs natively on NVIDIA's Blackwell cards, while older ones like the H200 cannot run it natively and move to FP8. If you do not have that hardware, you can try the model through providers: Baseten, Fireworks and Together, per Hugging Face.
The weights are open. The memory to hold them is not free.
The quality numbers are Reflection's and are not a comparison with other models. The table only shows how the three formats behave against each other; on SWEBench Verified BF16 is at 81.6%. So you will not read “beats so-and-so” here. Hugging Face also runs its own probe of five tasks and writes: “These are vibe evals, not benchmarks”. At low reasoning effort two of the five do not come out clean; at high, all five pass.
If you will host it on your own hardware, count from memory. Two B200s hold the NVFP4 version, but with almost no room for long context and several requests at once. That is why Hugging Face advises four, 720 GB in total.
Official primary sources
Hugging Face - Welcome Beam: a 501B open model you can own, customize, and scale, 16.10.2026Reflection - home page (Introducing Beam: Reflection's 501B open-weight model)The visual is generated code-art, no third-party images.