Reflection opens Beam: 501 billion parameters for coding

What this is about

Reflection AI introduced Beam on October 5, 2026, a large mixture-of-experts model with 501 billion total parameters and 23 billion parameters activated for each request. It focuses on coding and tasks in which a model uses tools and plans multiple steps. Reflection says it will publish the weights later in October. Until then, Beam is accessible through the company's platform but cannot yet be fully inspected locally.

The announcement matters because many of the strongest open-weight models have recently come from China. Beam is intended to provide a US-developed alternative that developers can run, adapt, and examine themselves. Whether it becomes genuinely open and practically useful will depend on the final license, the files that are released, and independent reproductions.

What Beam actually does

Beam uses a mixture-of-experts architecture. Put simply, the model contains many specialized subnetworks but activates only a small share of them for an individual token. Reflection says that 23 billion out of 501 billion parameters are active at a time. This can reduce computation compared with a dense model of a similar total size, but it does not remove the large storage requirement or the need for demanding infrastructure.

According to the company, Beam was pretrained on 23.8 trillion tokens. Training ran for four weeks on 10,500 Nvidia GB300 GPUs. Reflection identifies coding and multi-step agent tasks as its main areas. The company also claims three to four times higher inference efficiency than GLM-5.2. Vendor comparisons provide an initial reference, but they do not replace independent measurements made with the same hardware, software, and output quality.

Why it matters

Open weights give research teams and companies more control than a hosted-only API. They can operate a model in their own environment, study its behavior, add safeguards, and fine-tune it for specific work. This is particularly relevant when source code, customer data, or internal documents must remain inside an organization's infrastructure.

Beam therefore shifts the question from “Which API model is best?” to “Which organization can operate a very large model sensibly?” Twenty-three billion active parameters do not mean that only 23 billion parameters need to be stored. The complete weights, expert routing, and useful speed still demand substantial GPU memory. Smaller teams are more likely to use Beam through a provider or in heavily quantized form than on a single workstation.

In plain language

Beam is like an enormous workshop with hundreds of specialist departments. A dispatcher does not call every worker for every job; it sends each step to a few suitable teams. That can finish the job faster, but the entire building, all the tools, and all the specialists still have to exist.

This explains the difference between total and active parameters: only a portion works at once, while the much larger system still has to be stored and coordinated.

A practical example

A software company wants to run 2,000 automated code reviews every day. Beam could locate a defect, select tests, propose a patch, and check the change against project rules. Because 23 rather than 501 billion parameters are active for each token, computation may be lower than with an equally large dense model.

That number alone is not enough for a purchasing decision. The team would still need to measure throughput, latency, GPU memory, electricity cost, and the number of correctly solved internal tasks. If Beam gives strong suggestions across 2,000 reviews but requires an expensive multi-GPU installation, a smaller model may be the better economic choice.

Scope and limits

• The weights were not generally available at announcement time. The license, file set, and actual local usability need to be checked after release.
• Benchmark and efficiency claims initially come from the vendor. Independent tests may differ substantially with other hardware, quantization, or longer contexts.
• Open weights do not automatically make a model safe. Tool access, secrets, network permissions, and executed code still require sandboxing, logging, and human approvals.

Beam is therefore a serious candidate, but not yet a proven winner. The most useful test begins when outside teams can download the same weights and evaluate them under real conditions.

SEO & GEO keywords

Reflection AI, Beam, 501B, 23B active parameters, mixture of experts, open weights, coding model, AI agents, Nvidia GB300, local AI, model benchmarks