Ten days ago we wrote that Reflection's Beam was a 501B open model you couldn't download. That's no longer true. Since 16 October the weights sit in four public, ungated repos on Hugging Face, with a plain Apache 2.0 licence file. The catch is one layer down. Hugging Face's launch post says Beam has "day-0 support in vLLM", and on vLLM's GitHub that support is pull request #61978, opened the same morning and still unmerged.
The short answer
Beam 1.0 is 501.3B total and 23.4B active, text only, with a 1M-token context and a 30 June 2026 knowledge cutoff. The checkpoints weigh 1,002.68 GB in BF16, 503.30 GB in FP8 and 293.65 GB in NVFP4, plus a 3.86 GB DFlash2 drafter. Transformers 5.20.1, released today, loads it. The current releases of vLLM (0.31.0) and SGLang (0.5.22) don't: both projects have open pull requests, and SGLang's cookbook points to a dedicated Docker tag. The licence is unmodified Apache 2.0. The technical report is still marked "coming".
What "day-0 support" means today
Start with the repo, because it decides everything else. There's no modelling code in it. The config names model_type: beam and BeamForCausalLM, so trust_remote_code won't rescue you: your runtime has to know the architecture already. Transformers does, as of version 5.20.1, published on PyPI at 13:02 UTC on 16 October with Beam as its "special surprise".
vLLM doesn't, not in a release. Its latest tag is v0.31.0 from 5 October. The Beam work is PR #61978: 13,158 added lines across 85 files, opened at 07:36 UTC, with a first line that reads "Depends on FlashInfer nightly." It brings Beam's own reasoning and tool-call parsers, plus Blackwell kernels for the DFlash2 speculative decoder. SGLang looks the same. Version 0.5.22 went up on PyPI at 00:25 UTC, and native Beam support is PR #44747, opened at 13:40 UTC with about 27,000 added lines. The SGLang deployment cookbook, which did merge, runs every recipe from one image: lmsysorg/sglang:beam.
That's normal for a launch this size, and it isn't a scandal. It does change what "day-0" means for you. A plain pip install vllm today gets a build that can't load Beam. Would we put a vendor tag or an unmerged branch in production? Not us, not this week. A 13,000-line change that leans on a nightly dependency will move before it lands, and flags like VLLM_BEAM_ALWAYS_CONSTRAIN_OUTPUT might not survive review unchanged. Launch PRs can merge within days, mind you, and if this one does, the point's moot. If you just want to try the model, the Hugging Face post lists hosted access through Inference Providers (Baseten, Fireworks and Together).
The hardware, from the file sizes
We added up the file sizes from the Hugging Face API, in decimal gigabytes. The Hugging Face post says to "plan for four B200s (720 GB)" but doesn't say which checkpoint. Both NVFP4 and FP8 fit in that, and BF16 doesn't come close. It also calls NVFP4 "~280 GB"; the repo totals 293.65 GB. Small gap, but it's the one you'll feel when you size the KV cache.
NVFP4 is Blackwell only. SGLang's cookbook says so plainly and sends H200 owners to FP8 or BF16, both at tensor parallel 8 on one node. BF16 on eight H200s gives you 1,128 GB of HBM for 1,002.68 GB of weights, which leaves about 125 GB for the KV cache and everything else. It's a verified recipe, so it runs. We'd still want to see a long-context run on it before betting on anywhere near 1M tokens. On GB300, the cookbook needs two nodes for BF16.
What about eight H100s? 640 GB holds the 503 GB FP8 checkpoint on paper. Neither project publishes a recipe for that card, though, and we won't guess how much context would be left.
The licence is the plain one, the numbers are Reflection's
The LICENSE file is the stock Apache 2.0 text, appendix placeholders and all. The model card goes further than it needs to: Reflection "does not prohibit distillation, use of outputs to train other models", along with fine-tuning and merging, and it says the card doesn't restrict the licence. Compare that with Meta's Muse Spark Open licence, published this morning, which wants a credit line on models trained on its outputs. If you're building a fine-tune or a distilled model for clients, Beam is the simpler paper.
The benchmarks are another matter. Every score on the card is self-reported, sampled at temperature 0.7, with no technical report behind it yet. SWE-bench Verified reads 81.6% in BF16 and 80.7% in NVFP4. Most rows hold within about a point across the three precisions, SciCode slips from 49.2% to 46.9%, and SWE Bench Pro v2-Hard goes the other way (76.5% in BF16, 78.6% in NVFP4), which tells you how much run-to-run noise sits in those numbers. The claim we'd most like to see backed is the headline one, roughly 3x fewer reasoning tokens than leading open models. There's no table for it on the card. I might be wrong to wait for the report, but that's the number that decides what Beam costs to serve.
One loose end: Reflection's own announcement page still says the weights arrive "later this month" and offers early-access sign-up. The repos say otherwise. We've moved Beam from promised to published on our open-weight models tracker, next to Mistral Large 4, which is still waiting on its weights.
FAQ
Are Reflection Beam's weights available to download?
Yes, since 16 October 2026. Four public, ungated repos under the reflection organisation on Hugging Face hold the BF16 (1,002.68 GB), FP8 (503.30 GB) and NVFP4 (293.65 GB) checkpoints and a 3.86 GB DFlash2 drafter, all under Apache 2.0.
Can I serve Beam with the current vLLM release?
No. vLLM v0.31.0 doesn't include the Beam architecture, and support is pull request #61978, opened on 16 October and not merged when we checked. It depends on a FlashInfer nightly. SGLang's support is also an open pull request, and its cookbook uses the lmsysorg/sglang:beam Docker image.
What GPUs does Beam need?
The NVFP4 checkpoint needs Blackwell; Hugging Face suggests four B200s with 720 GB in total. On Hopper, SGLang's cookbook has FP8 and BF16 recipes for a single node of eight H200s. BF16 on GB300 takes two nodes.
Can I train another model on Beam's outputs?
Yes. The licence is unmodified Apache 2.0, and the model card says Reflection doesn't prohibit distillation or using outputs to train other models. Its usage policy is linked as a recommendation, and the card says it doesn't restrict the licence.
Sources: Beam model card and licence, Hugging Face model API, Hugging Face launch post, vLLM PR #61978, SGLang PR #44747, SGLang Beam cookbook, Transformers v5.20.1 release, Reflection announcement. Our earlier write-up: Reflection Beam is a 501B open model with no weights yet.
