Reflection AI unveils Beam, its first open-weight model

Reflection AI finally has a model to show. Not weights. Not a benchmark you can rerun. A date.

The lab introduced Beam on Monday, its first open-weight model. Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agent work. Reflection says the weights, a technical report, a model card and developer tools will follow later this month under the Apache 2.0 license.

What Reflection says Beam is

Reflection says it pretrained Beam on 23.8 trillion tokens drawn from the web, public sources and licensed datasets, with an emphasis on source code, technical explanations and math and science material.

Pretraining ran on 6,144 NVIDIA GB300 GPUs and finished in under four weeks, the company says. Reinforcement learning then used about 10,500 GB300 GPUs over another four weeks. Reflection says it extended the context length to 1 million tokens during midtraining.

The pitch is efficiency. Reflection says Beam is competitive with GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks. On advanced reasoning benchmarks, it says Beam matches GLM 5.2's scores while using 3 to 4 times less inference compute.

3 to 4 times less compute.

Beam is still in final red-teaming and evaluation. Early access runs through a waitlist on Reflection's platform.

An American answer, on paper

Look at the comparison Reflection chose. GLM and Qwen, the two models it measures Beam against, both come from Chinese labs. Reflection frames Beam as a step forward for the Western open-weight frontier.

Beam is that pitch made concrete. A model with 23 billion active parameters is cheaper to serve than its total size suggests, and Apache 2.0 is about as permissive as licenses get. If the efficiency claim holds, Beam could be the open model a US company picks when its legal team will not sign off on a Chinese one.

That is a big if. Today, every number above comes from Reflection.

What it means for you

If you run open models in production, sign up for the waitlist and plan an eval for the week the weights land. Test it on your own coding and agent tasks against whatever you run now.

Judge it on cost per finished task. That is where the efficiency claim either shows up or does not.

If you are choosing an open model this month, you do not need to wait for Beam. You do need to know it is coming.

What to watch

The release of the weights and the technical report later this month is the test. Watch for independent benchmark runs within days of it, and for which inference providers host Beam first.

Weights, then verdict.