Reflection’s first model is now something you can actually download. On October 16, 2026, the AI lab put Beam’s weights on Hugging Face under an Apache 2.0 license: a 501-billion-parameter mixture-of-experts model with 23 billion parameters active per token and a 1-million-token context window. The weights are free. Serving them yourself, by Reflection’s own deployment docs, means four NVIDIA B300 GPUs wired together with NVLink.
That gap between “free” and “runnable” is the real story for most teams. Beam arrives with a permissive license, three precision formats and day-one support across more than 40 partners. But Reflection hasn’t published a per-token price for its own API, the technical report is still marked “coming,” and every benchmark number so far comes from Reflection.
What Reflection actually shipped
The Hugging Face model card lists 501.3 billion total parameters and 23.4 billion active, across 52 layers (3 dense, 49 mixture-of-experts) with 256 experts. Hugging Face’s launch post adds that 8 experts are routed per token. It takes text in and puts text out, with no image or video input, and Reflection says it was developed predominantly for English with a knowledge cutoff of June 30, 2026.
Reflection released three versions of the instruction-tuned model, plus a separate “DFlash2” drafter model for speculative decoding (a trick that drafts tokens with a small model so the big one generates faster). The checkpoint sizes, from Reflection’s docs:
| Checkpoint | Download size | Hugging Face repo |
|---|---|---|
| BF16 (full precision) | 933.81 GiB | reflection/Beam-1.0-501B-A23B |
| FP8 | 468.72 GiB | reflection/Beam-1.0-501B-A23B-FP8 |
| NVFP4 (4-bit) | 273 GiB | reflection/Beam-1.0-501B-A23B-NVFP4 |
The license is the headline for anyone building products. Apache 2.0 allows commercial use, and the model card goes further than most: Reflection says it does not prohibit distillation, using Beam’s outputs to train other models, fine-tuning, quantization or merging.
Beam was first announced on October 5 with an early-access sign-up; that introduction post promised the weights “later this month.” Reflection delivered 11 days later.
Free weights, four B300s to serve them
Reflection’s serving guides for SGLang, vLLM and TokenSpeed all use the 4-bit NVFP4 checkpoint, and the listed requirement is “four NVIDIA B300 SXM GPUs connected by NVLink, on a Linux amd64 host.” Those are NVIDIA’s current Blackwell Ultra data-center parts, not something you put under a desk. Reflection doesn’t publish a smaller supported configuration, and it doesn’t list one for the FP8 or BF16 checkpoints, which are about 1.7 and 3.4 times bigger on disk.
So “open” here means control, not cheap local inference. A company can run Beam inside its own cloud account or data center, keep prompts and code on its own hardware, and post-train it. Reflection says Dell and HPE will deliver Beam for on-premises and air-gapped deployments. Hobbyists with a gaming GPU are not the audience, whatever the LM Studio and Ollama names on the partner list suggest.
If that hardware bill sounds familiar, it’s the same tension we flagged when Meta opened Muse Spark for download: the weights are free, and the GPUs are where the money goes.
What Beam costs if you rent it
Reflection’s own API is in public beta, and the company says it’s “free to start” at platform.reflection.ai. It hasn’t published a per-million-token rate card in its launch post or docs that we could find. The hosted market has set prices anyway. OpenRouter’s live endpoint list for Beam, checked the morning of October 17, showed four providers:
| Provider (via OpenRouter) | Input / 1M tokens | Output / 1M tokens | Precision listed |
|---|---|---|---|
| Nebius | $0.25 | $1.00 | FP8 |
| Together | $0.40 | $1.20 | not stated |
| CoreWeave | $1.00 | $3.00 | FP4 |
| Baseten | $1.30 | $4.30 | FP4 |
That’s a more than fourfold spread on input price for the same weights, and the cheapest listing runs the higher-precision FP8 checkpoint. Expect those numbers to move as more of the launch partners (AWS, Google Cloud, Microsoft Foundry, Fireworks and others named by Reflection) bring endpoints online. For another recent open-weights flagship and what it costs to use, see our Mistral Large 4 breakdown.
The benchmarks are Reflection’s, and the report isn’t out
Reflection’s model card reports Beam’s scores across all three precisions, run at temperature 0.7. For the BF16 model, Reflection lists 81.6% on SWE-bench Verified, 80.1% on Terminal Bench v2.1, 66.1% on SWE Bench Pro v1, 97.7% on AIME 2026, 90.6% on GPQA Diamond and 88.3% on LiveCodeBench. On Humanity’s Last Exam it reports 36.0% without tools and 51.1% with them. The 4-bit NVFP4 version lands within about a point on most coding tests (80.7% on SWE-bench Verified, for example), which, if it holds up, is the reason that checkpoint is the default for serving.
Two caveats. The model card’s table only shows Beam; it doesn’t put rival models next to it, so any “beats X” framing you see elsewhere comes from Reflection’s earlier blog charts or third parties. And the technical report is still listed as “coming.” Until it lands, and until independent leaderboards rerun these tests, treat the numbers as Reflection’s.
The efficiency pitch is the more interesting claim. The model card says Beam reaches comparable quality to leading open models using about three times fewer reasoning tokens. Google Cloud, a launch partner, puts it at “3x to 4x” in its day-zero post, where Reflection’s head of open source, Joe Spisak, says deploying on Google Cloud delivered “up to 5x higher performance and 8x lower latency.” Fewer tokens per answer is a direct cost cut on any per-token bill, so this is the claim to test first.
Not a robot brain, but robotics teams will feel it
Beam is text-only. It can’t look through a camera, so it isn’t a vision-language-action model and it won’t drive a humanoid. (The “plans a robot’s escape” demo in Reflection’s launch post is a puzzle game played through tool calls.) Where it matters for robotics is the software around the robot: writing and fixing code, running long agent loops, and doing that on hardware a lab or manufacturer controls, under a license that permits fine-tuning on private data.
Reflection also published safety numbers from its in-house Iris evaluations: an 88.9% refusal-or-redirect rate on harmful requests and 81.5% compliance on harmless ones for the BF16 model. That second figure means roughly one in five benign prompts in its test set got pushback, something teams building agents will notice. The full methodology is in Reflection’s separate safety report.
Reflection, which Reuters reported raised $2 billion at an $8 billion valuation in October 2025 with NVIDIA’s backing, has now shipped the thing that money was meant to buy. The next checkpoints are concrete: the technical report, a published price for Reflection’s own API, and the self-distillation recipes the company says it will release “in the coming days.”
Frequently asked questions
What is Reflection Beam?
Beam is Reflection’s first open-weight language model, released October 16, 2026 on Hugging Face under Apache 2.0. It’s a sparse mixture-of-experts model with 501.3 billion total and 23.4 billion active parameters, a 1-million-token context window and text-only input and output, aimed at coding, reasoning and agent tasks.
How much does Reflection Beam cost?
The weights are free to download. Reflection’s own API is a public beta it describes as free to start, with no per-token rate published in its launch materials. Hosted prices on OpenRouter on October 17 ranged from $0.25 input and $1.00 output per million tokens (Nebius) to $1.30 and $4.30 (Baseten).
What hardware do you need to run Beam yourself?
Reflection’s serving guides list four NVIDIA B300 SXM GPUs connected by NVLink on a Linux amd64 host, using the 4-bit NVFP4 checkpoint (273 GiB). The FP8 checkpoint is 468.72 GiB and the BF16 checkpoint is 933.81 GiB. Reflection hasn’t published a smaller supported setup.
Can I use Beam commercially or fine-tune it?
Yes. Apache 2.0 permits commercial use, and Reflection’s model card says it does not prohibit fine-tuning, quantization, merging, distillation or training other models on Beam’s outputs. Reflection points users to its usage policy for areas to review.
How good is Beam on benchmarks?
By Reflection’s own model card, the BF16 version scores 81.6% on SWE-bench Verified, 80.1% on Terminal Bench v2.1, 97.7% on AIME 2026 and 90.6% on GPQA Diamond. Those are company-reported results; the technical report hasn’t been published yet and independent reruns are still pending.
Can Beam control robots?
No. Beam is text-only, so it can’t process camera input and isn’t a vision-language-action model. Robotics teams could use it for code generation, tooling and agent workflows running on their own hardware.
Sources
- Reflection, “Beam is now available,” October 16, 2026: reflection.ai
- Reflection, “Introducing Beam: Reflection’s 501B open-weight model,” October 5, 2026 (updated October 8): reflection.ai
- Reflection, Beam-1.0-501B-A23B model card (architecture, benchmarks, safety evaluations, usage), Hugging Face: huggingface.co
- Reflection Developer Docs, “Download the weights” (GPU requirements, checkpoint sizes), last modified October 16, 2026: developers.reflection.ai
- Hugging Face, “Welcome Beam: a 501B open model you can own, customize, and scale,” October 16, 2026: huggingface.co
- Google Cloud, “Day-0 support with 5x perf for Reflection Beam on Google Cloud,” October 16, 2026: cloud.google.com
- OpenRouter, Beam 1.0 501B A23B model and provider endpoint listing (prices checked October 17, 2026): openrouter.ai
- Reuters, “Nvidia-backed Reflection AI raises $2 billion in funding, boosts valuation to $8 billion,” October 9, 2025: reuters.com
All sources accessed October 17, 2026. OpenRouter provider prices checked the morning of October 17, 2026, and may change.
Related: Muse Spark open download: price and hardware · Mistral Large 4 price and open weights · More LLM News
Last updated: October 17, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
