What is Reflection's Beam, and why does it matter that it is Apache 2.0?

Yesterday, Reflection AI published the weights of Beam on Hugging Face under the Apache 2.0 licence. Weights are the learned numbers that make up a model, so anyone can now download and run it. The company had announced the model on October 5 and promised the files by the end of the month.

Beam is not the biggest or the highest-scoring open model. Reflection itself says Kimi K3 remains ahead on raw capability. What it offers instead is a clean licence from an American lab, plus a claim about efficiency that buyers can now test.

What was released

A parameter is one learned number, and a token is a small piece of text. Beam is a mixture-of-experts model, in which only a portion of it works on each token. The three formats store each number in fewer or more bits: 16, 8 or about 4. That is quantization, and it explains why the same model has three very different download sizes.

Text-only is a real limit. TechCrunch notes that Thinking Machines' Inkling, perhaps its most direct US rival, accepts more kinds of input. If your work depends on screenshots, scanned forms or charts, Beam is not your model.

How it was trained

Reflection's post puts the weight on reinforcement learning (RL), a stage where a model attempts tasks, gets graded and adjusts. Its numbers are large: more than 100 million attempts, called rollouts, on 10,500 NVIDIA GB300 chips over four weeks, with about 1.3 billion sandboxes (isolated environments for running and grading code), and a million environments for coding, agent and science tasks.

The hard part it describes is staleness. When a run lasts a long time, the early part of an attempt was written by an older version of the model than the one now being trained. Imagine a coach correcting a player from a game film recorded the day before, when the player has already changed technique. Reflection says its methods keep learning stable even with interactions more than a day old.

The firm's main claim is efficiency: scores comparable with GLM-5.2 on reasoning tests while using three to four times less inference compute. It estimated that compute from active parameters and tokens generated, which leaves out prompt reading and other costs, as its own notes say. TechCrunch stresses that the performance claims have not been independently verified.

Reflection's own table shows it behind the top open models in raw score. Note that it publishes different DeepSWE numbers for rivals than those labs report for themselves, which shows how much the test setup matters.

What it takes to run

The sums are short. The NVFP4 files are about 294 GB, which means at least four 80 GB graphics cards for the weights, plus working memory for long context. So this is still server hardware, though well below the largest open models. Reflection says Beam has day-one support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA's Dynamo, and comes with a drafter model for speculative decoding, where a small model proposes text that the main one checks to save time.

It also launched Mirror CLI, a terminal agent that lets developers inspect activity and approve tool calls, and says more than 40 partners and integrations are involved.

Why Apache 2.0 matters

Apache 2.0 has no revenue threshold, no attribution banner on your product and no separate deal for hosting. It asks you to keep the licence and notices when you redistribute and to mark changed files, and it includes an express patent licence from contributors. You can check this in the licence file.

That differs from the custom licences on many large open models, which add conditions for model-hosting businesses or require attribution. Reflection does publish a separate usage policy, and its card asks developers to check legal requirements such as privacy and export rules. It's worth reading, even if the licence itself is short.

Who should choose Beam over a larger open model?

Three groups have a real reason to look at it. Hosting companies avoid the separate agreements that some rival licences demand. Organizations with rules about Chinese suppliers, such as some public bodies and defence firms, get a large model from a US lab, trained on data that Reflection describes as coming from the web and licensed datasets. And teams paying per token may save money if the efficiency claim holds on their tasks.

Reflection aims at governments and enterprises as well. TechCrunch reports its "AI factory" plan, where institutions train its models on their own private data, and a test with South Korea's Shinsegae Group. A Canadian public body weighing sovereignty should still note that an American supplier raises its own jurisdiction questions, although running the weights on your own machines removes the supplier from the data path.

If you need the top scores, image input or strong French out of the box, look elsewhere first. The card says Beam was developed predominantly for English, so test French yourself before relying on it in Québec.