Reflection AI unveils Beam, a 501B-parameter open model due this month

Reflection AI has unveiled Beam, an open-weight AI model with 501 billion total parameters. The full release is expected later this month.

The company announced the model on October 5, 2026. It is pitching Beam as a Western answer to the Chinese open models that have quietly taken over a big slice of developer workloads.

What Beam actually is

The technical name for this design is a sparse Mixture-of-Experts architecture. Beam has 501 billion parameters in total, but only 23 billion are active for any given token. A token is roughly a word or a chunk of a word.

That gap matters because active parameters drive the cost of running a model. A huge model that only uses a small fraction of itself at a time can stay smart without burning cash on every reply.

Reflection says Beam was pretrained on 23.8 trillion tokens. It then went through large-scale reinforcement learning, a process where the model practices tasks and gets rewarded for good answers. The company ran more than 100 million of these practice rollouts on 10.5K NVIDIA GB300 GPUs.

The benchmark pitch

Reflection is positioning Beam against two specific Chinese rivals: Z.ai’s GLM-5.2 and Alibaba’s Qwen 3.8-Max. According to the company’s claims, Beam matches or approaches those models on performance, especially in reasoning and coding.

Beam posted a score of 80.9 on SWE-Bench Verified, which tests whether a model can fix real software bugs. It scored 80.1 on Terminal Bench v2.1, which measures how well a model operates in a command-line environment.

Reflection says Beam reaches that competitive level while using 3 to 4 times less inference compute.

Benchmarks are self-reported at this stage, and the technical report has not yet been published. Independent testing will start once the weights are out.

The full model weights and technical documentation are slated for release later in October 2026 under the Apache 2.0 license. Apache 2.0 is one of the most permissive open-source licenses around. It generally lets companies use, modify and commercialize the software with few strings attached.

Who is behind it

Reflection AI is a US startup founded in March 2024 by former Google DeepMind researchers.

The company has raised more than $4 billion since its founding. Its latest funding round put its pre-money valuation at $25 billion. Backers include Nvidia, Sequoia and Citigroup.

Reflection has also built ties with Washington. Its partnerships include work with the Pentagon and the US Department of Energy. The company says its focus is shifting toward scalable AI tailored to enterprise customers.

Why the China angle matters

The open-model market has a geography problem, at least from a US perspective. Chinese open models have reportedly captured more than 30% of token share recently.

Reflection is betting there is real demand for a homegrown alternative. Its stated aim is to give Western users robust open-weight options that do not depend on Chinese providers.

For government agencies and regulated industries, where a model comes from can matter as much as how well it performs. Reflection’s work with the Pentagon and Department of Energy suggests it is leaning hard into that argument.

What this means

For enterprise buyers, the efficiency claim is the story to watch. If Beam really delivers comparable reasoning and coding performance with 3 to 4 times less inference compute, it changes the math on self-hosting.

There are clear risks. Self-reported benchmarks have a long history of looking better in a launch post than in the wild. Coding scores in particular can be sensitive to how tests are run.

The key dates to watch are the planned weight release and the technical report later in October 2026.