Reflection Beam: 501B Open Model, Apache 2.0
Reflection released Beam, a 501B-parameter open-weight model for coding and agents, with Apache 2.0 licensing and a 1M-token context. Here is what it takes to use it.
Reflection has released Beam, an open-weight, text-only language model for coding and agent work, and the weights are under the Apache 2.0 license. It has 501 billion parameters in total, 23 billion active per token, and a 1 million token context window. The model page went up on Hugging Face on 16 October 2026, and it calls Beam the first model from Reflection, a new US lab.
The headline is the license and the size. The part most builders will care about is the memory bill, because a mixture-of-experts model is cheap to run per token and expensive to hold.
What Reflection published about Beam
Everything below comes from the Hugging Face model page, which is the only primary source for the release.
Architecture: a mixture-of-experts transformer with 52 layers, 256 experts per MoE layer and 8 experts routed per token. Hidden size is 5,120.
Attention: 43 sliding-window layers with a 512-token window, plus 9 global layers.
Context: 1 million tokens, text in and text out.
License: Apache 2.0, with no additional terms listed on the page.
Checkpoints: BF16 (about 1 TB), FP8, and NVFP4 (about 280 GB, aimed at Blackwell GPUs).
Software support: transformers, vLLM and SGLang.
If the mixture-of-experts idea is new to you, our plain guide to mixture of experts covers why only a slice of the model runs for each token.
The memory arithmetic the page does not spell out
Active parameters decide how much computation each token costs. Total parameters decide how much memory you need before the first token. For Beam those are 23 billion and 501 billion, a gap of more than 21 times.
Here is the weights-only math per checkpoint, using the bits per weight for each format. The NVFP4 figure uses 4.5 bits per value, the effective storage NVIDIA describes in its NVFP4 write-up, which counts the shared scale factors.
Checkpoint | Bits per weight | Weights alone (our arithmetic) | Hardware the page names |
|---|---|---|---|
BF16 | 16 | about 1,002 GB | not listed |
FP8 | 8 | about 501 GB | 8 H200s in the SGLang example |
NVFP4 | 4.5 | about 282 GB | 4 B200s (720 GB), or 2 B200s (360 GB) with little room left |
Those are weights only. The KV cache and activations come on top, and a 1 million token context makes the cache a real cost. That is why the page warns that two B200s can hold the NVFP4 weights but leave little headroom.
For scale, the VRAM guide for local models shows where consumer cards top out. Beam is not in that range. Even the smallest checkpoint needs data-center GPUs, so "open weights" here means a team with its own cluster, not a laptop project.
The benchmark numbers, and how to read them
Reflection reports these results for the BF16, FP8 and NVFP4 checkpoints in that order:
Benchmark | BF16 | FP8 | NVFP4 |
|---|---|---|---|
SWE-bench Verified | 81.6% | 80.2% | 80.7% |
Terminal Bench v2.1 | 80.1% | 78.9% | 78.2% |
LiveCodeBench | 88.3% | 88.1% | 87.4% |
GPQA Diamond | 90.6% | 90.3% | 89.3% |
Two things stand out. First, the lab's own table shows the 4-bit checkpoint staying within about two points of BF16 on every row above, which is the case for quantization working well on a model this size. Second, these are the lab's own numbers, published on its own page. Treat them as a claim to test, not a ranking.
What a small team should actually do
Most founders and freelancers will not host a 501B model. The realistic options are the ones the page lists: Beam is available through Hugging Face Inference Providers (Baseten, Fireworks and Together are named) and can be deployed with one click on Inference Endpoints. The page does not state pricing for any of them, so check each provider before you plan around it.
If you want to evaluate Beam for a coding workflow, run it against ten real tasks from your own repository before you believe any public score. The license is the easy part here. On that front the page is clear, and it contrasts with the conditions attached to Meta's release this week, which we covered in our Muse Spark Open breakdown. For the wider picture of what each license family lets you ship, see AI model licenses explained.
To keep track of releases like this without reading every model page, our guide to keeping up with AI news lays out a routine that takes minutes a day.
Beam FAQ
Can I run Reflection Beam on my own computer?
Not on a laptop or a single consumer GPU. The smallest checkpoint is about 280 GB for the weights, and the page suggests planning for four B200 GPUs for the NVFP4 version.
Is Beam open source?
The weights are released under Apache 2.0, a permissive license. The page describes Beam as open-weight, and it does not say whether training data or training code is also released.
What does 23B active parameters mean for Beam?
For each token, the model routes through 8 of 256 experts per layer, so about 23 billion of its 501 billion parameters do the work. That lowers compute per token, but all 501 billion still have to sit in memory.
Where can I use Beam without hosting it?
The model page names Baseten, Fireworks and Together through Hugging Face Inference Providers, plus one-click deployment on Inference Endpoints. Pricing is not stated on the page.
How did this land?
About the author

Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.


