The Reflection AI Beam weights are now public. On October 16, Reflection released the 501-billion-parameter coding model on Hugging Face under the Apache 2.0 license, so you can download it, run it, and even sell products built on it.
That is a big shift from the October 5 announcement, when Beam was only a waitlist and a promise. Here is what you actually get, who is hosting it, and what is still missing.
The short version, if you are skimming:
- Released: October 16, 2026, on Hugging Face (huggingface.co/reflection)
- Size: 501B total parameters, 23B active per token (sparse Mixture-of-Experts)
- License: Apache 2.0, so commercial use is allowed
- Formats: BF16, FP8 and NVFP4, plus a small drafter model for faster decoding
- API price: not published; free to start during the public beta
Those five lines cover most of what a developer needs to decide whether to try it this week.
What Reflection actually released
According to Reflection’s release post, the drop includes the model weights, a model card with final benchmarks, and a safety technical report. You also get three quantized versions, which matter because a 501B model is huge in full precision.
One thing is still coming. Reflection says self-distillation recipes, meaning code to shrink Beam into a smaller model of your own, will arrive “in the coming days.” SiliconANGLE also reports fine-tuning templates, evaluation tools, and ready-made serving configs.
The “mixture-of-experts” design is why the 23B active figure matters. Only a slice of the model runs for each token, so speed and cost behave more like a mid-size model than a 501B one. You still need memory for all the weights, though.
How strong is Beam at coding?
Beam is built for coding and agent work, and the scores Reflection published reflect that. These come from the company’s own announcement post, so treat them as vendor claims until independent tests land.
| Benchmark | Beam score |
|---|---|
| SWE-Bench Verified | 80.9 |
| SWE-Bench Pro v2-Hard | 77.2 |
| SWE-Bench Pro v1 | 65.5 |
| SWE-Bench Multilingual | 78.0 |
| Terminal-Bench v2.1 | 80.1 |
An 80.9 on SWE-Bench Verified puts Beam in the same conversation as top closed models for fixing real GitHub issues. Still, benchmark numbers rarely match how a model feels in your own codebase, so test it on your own repos before switching anything.
If you follow open models, this lands in the same space as Meta’s Muse Spark Open and Mistral Large 4. The difference is the clean Apache 2.0 license and the heavy coding focus.
Where you can run the Reflection AI Beam weights
Reflection lists a long set of hosting partners, so most people will not need to host it themselves. These are the main routes:
- Cloud platforms: AWS, Google Cloud, Microsoft Azure, Oracle, CoreWeave, Nebius, Nscale and Databricks
- Inference providers: Together AI, Fireworks AI, Baseten, Modal, Cloudflare and OpenRouter
- Local and private setups: Ollama, LM Studio, Nvidia NIMs, and Dell or HPE for air-gapped installs
If your team wants code to stay inside your own network, the Dell and HPE options are the interesting ones. Everyone else can start with a hosted endpoint and avoid the hardware bill.
Reflection has not published hardware requirements. The model was trained on NVIDIA GB300 GPUs, and its own production inference runs on Blackwell hardware, which hints that the full-size version is not a laptop job.
What is still unconfirmed
Three gaps are worth knowing before you plan around Beam. Reflection has not stated per-token API pricing, only that the public beta is free to start. The context length for the released weights is not stated either, though the earlier announcement said mid-training stretched effective context to 1 million tokens.
The third gap is independent testing. SiliconANGLE notes the claim that Beam approaches models four times its size is not independently verified, and the quantized versions trade some quality for efficiency. Give the community a week or two of third-party evals before trusting the headline numbers.
Frequently Asked Questions
Is Beam really open source?
The weights are open under Apache 2.0, which allows commercial use. That makes it “open-weight.” Reflection has not said the full training data or code is public.
Where do I download Beam?
From Reflection’s Hugging Face page at huggingface.co/reflection. You can pick BF16, FP8 or NVFP4 files.
Can I run Beam on a normal PC?
Probably not the full model. Reflection has not listed requirements, and 501B parameters need a lot of GPU memory even when quantized. A hosted API is the easier start.
How much does the Beam API cost?
Reflection has not published rates. It says Beam is free to start on its platform during the public beta.
My take: the Apache 2.0 license and the wide hosting list make Beam worth a weekend test if you write code with AI agents. Run it on a real repo through a hosted endpoint first, compare it to what you use now, and wait for independent benchmarks before moving any production work.
Leave a Reply