New US lab gives away a giant AI free

Reflection AI, a new US lab, has released the full weights of its first big model, called Beam. Anyone can download it, and the license is a very open one. That matters because a company can run a model this size on its own machines.

What happened

On 16 October 2026, Reflection published Beam on Hugging Face, a site where people share AI models. The model is named Beam-1.0-501B-A23B, and the lab says it is built for long documents and for "agent" work, where an AI takes many steps on its own. According to Hugging Face, it reads and writes text only.

The license is Apache 2.0, a well-known open license. There is no price tag on the download.

Reflection did not stop at the model. Hugging Face says the lab also shared a helper model that makes answers come out faster, plus recipes for teaching Beam new skills. SiliconANGLE adds that the lab plans more recipes in the coming days.

Key facts

• Size: 501 billion parameters in total, with 23 billion active for each word it handles, according to Hugging Face. Parameters are the numbers inside a model that hold what it learned.
• Memory: it can take in 1 million tokens at once (a token is a small piece of a word), as listed on the same Hugging Face post.
• Coding score: 81.6% on SWE-Bench Verified for the full-size version, a test where the AI fixes real software bugs. Reflection supplied the score.
• Other scores supplied by Reflection: 80.1% on Terminal Bench v2.1 and 90.6% on the GPQA Diamond science test.
• Training: about two months on several thousand high-end graphics chips, says SiliconANGLE.
• Smaller versions: Reflection also released two shrunken copies, called FP8 and NVFP4, that need less memory and score a little lower.

Background

Beam is the first model from this lab. SiliconANGLE says it was first introduced about a week before the weights came out, around 5 October.

The word "active" is the key to the size. Beam is a "mixture of experts" model. Hugging Face says each layer has 256 small expert parts, and only 8 are used for each token. So the model is huge, but only a slice of it works at any moment. SiliconANGLE says the lab claims it gets close to models four times its size on some tests, while using far less computing power and memory.

What it means for you

If you run a company, you can now host a top-size AI on your own machines. Your data stays with you. You do not pay a per-use fee to another company.

If you are not a developer, you will not run Beam at home. The smaller NVFP4 version still needs real hardware. Hugging Face suggests four B200 chips as one setup, and those are data center chips. But you may use Beam without knowing it, because the post lists Baseten, Fireworks and Together as services that host it.

Our take: the test scores come from the lab itself. Wait for outside testers to repeat them before you trust the claims.

What happens next

Reflection says more training recipes are coming. Hugging Face also says the lab will give $10,000 in credits to the best projects built with Beam.

Our take: watch for independent test results over the next few weeks. They will show if Beam really matches the lab's numbers.

FAQ

What is Beam?

Beam is a very large open AI model from a new US lab called Reflection AI. It is made for coding and long, multi-step tasks, and it was released on 16 October 2026.

Can I use Beam for free in my business?

Our take: very likely yes. Hugging Face lists the license as Apache 2.0, which is an open license, but read it yourself before you build on it. You still pay for the computers needed to run it, or for a host.

Can I run Beam on my own laptop?

Almost certainly not. The model is huge, and the post lists data center chips as the recommended hardware.

Sources

• huggingface.co
• siliconangle.com