Reflection AI launches Beam, 501 billion open model parameters that challenges China with less computational power

Reflection AI, an American artificial intelligence startup, presented Beam, an open weight model with 501 billion parameters aimed at coding and agent tasks. According to the company, performance is comparable to that of GLM-5.2 in China, consuming from one third to one quarter of the computational power required to train equivalent models. It is the first open model of a US startup positioned to compete from equal to equal with the main Chinese LLMs.

Beam: What Reflection AI launched with 501 billion parameters

Beam is not a general purpose chatbot. The company describes it as a model specialized in two uses: writing and reviewing code, and operating as an agent — that is, executing sequences of actions with little human supervision, rather than just answering isolated questions.

The number that startup highlights is the size: 501 billion parameters, approximate measurement of the capacity of a model. More relevant for those who want to adopt the technology is the announced license: the full weights will be released under Apache 2.0, which allows commercial use, modification and redistribution without royalties, including execution on own infrastructure.

Why computational spending turned the battlefield into AI

According to Reflection AI itself, Beam achieves performance comparable to GLM-5.2 using between a third and a quarter of the computational power. If the number is supported by external evaluations, it changes the industry's economy: training and serving border models is expensive, and the energy and GPU account usually decides which companies can compete.

The statement, however, is from the company that launched the model — not from an independent benchmark. It is the kind of data that needs to be confirmed by third parties before it is treated as a technical consensus.

How Beam compares to Chinese models like GLM-5.2

Zhipu's GLM-5.2 had already drawn attention by rivaling American models charging a fraction of the price — the portal covered this movement in Zhipu GLM 5.2: Chinese model rivals American AI for a fifth of the price.

Beam enters a board already occupied. Moonshot AI released the Kimi K3, with 3 trillion parameters, and DeepSeek presented the 1.6 trillion V4. Nvidia also attempted the reverse movement, with Nemotron 3 Ultra, of 550 billion parameters, without losing Chinese leadership in open models. The differential that Reflection AI claims is not the raw size, but the efficiency per parameter.

What Beam Changes to Devs and Companies in Brazil

An Apache 2.0 license on full weights means that Brazilian companies can download Beam, adjust it with their own data and run it on local or cloud servers hired in the country — something relevant for sectors with data residence restrictions and for those who need to demonstrate control over sensitive information under LGPD.

In practice, the cost of cloud inference is still quoted in dollar, which keeps the account sensitive to exchange. The real advantage appears in projects that today depend on subscription-closed APIs that could migrate to a self-hosted model, provided they have hardware suitable for such a model.

What Beam's announcement hasn't solved yet

There are so far no independent benchmark results, architecture details, context window size, volume of training data or minimum hardware requirements publicly disclosed. Nor is it clear the exact schedule of release of weights or whether there will be minor variants of Beam.

For those who decide to adopt technology, the recommendation is to expect the effective publication of weights and the first tests of the community before replacing models in production. The original source of the ad is TechCrunch report.

Frequently Asked Questions about Reflection AI Beam

What is Beam from Reflection AI?

It is an open weight AI model with 501 billion parameters, created by the American startup Reflection AI and focused on coding and agent tasks. The full weights will be released under Apache 2.0 license, according to the company.

Is Beam better than Chinese models?

Reflection AI states that performance is comparable to GLM-5.2 using one-third to one-quarter of computational power. As the comparison part of the company itself and there are still no independent benchmarks, the data should be treated as a declaration by the manufacturer.

Can I use Beam commercially in Brazil?

The Apache 2.0 license, announced for full weights, allows commercial use and modification without royalties. Still, it is necessary to wait for the effective release of weights and verify hardware requirements to run a model of 501 billion parameters.