Reflection AI just introduced Beam as its first open-weight model, putting it forward for coding, reasoning, and agentic workloads. The announcement did more than outline a new model: it named Z.ai’s GLM-5.2, Alibaba’s Qwen 3.8-Max, and Kimi K3 as the points of comparison. Reflection said Beam remained in final red-teaming and evaluation. Its launch announcement said the weights, technical report, model card, and developer artifacts would arrive later in October.
The timing gives the release a provisional quality. Reflection has published an account of Beam’s scale, training run, and comparative results before putting the associated materials into developers’ hands. Reuters identified the company’s founders, Misha Laskin and Ioannis Antonoglou, as former DeepMind researchers. The Oct. 5 announcement is their company’s first public open-weight-model launch, with the promised materials still to come.
Beam’s Size and Training Account
Reflection says Beam has 501 billion total parameters, with 23 billion active. Those figures anchor the company’s description of a model built for coding, reasoning, and agentic workloads. A total-parameter count and an active-parameter count describe different parts of the model’s stated configuration, and Reflection presented both alongside its task focus rather than as a score for every possible use.
The company also set out an unusually detailed training account. It says Beam was pretrained on 23.8 trillion curated tokens. For a subsequent reinforcement-learning campaign, Reflection reported using 10,500 NVIDIA GB300 GPUs over four weeks, producing more than 100 million rollouts and using about 1.3 billion sandboxes. The figures describe the training work Reflection says preceded its planned release of the model’s weights.
That sequence matters because Beam’s Oct. 5 debut was an announcement, not the release of every accompanying file. Reflection said it was finishing red-teaming and evaluation, then planned to publish the weights along with a technical report, model card, and developer materials later in the month. Until then, the public account consists of the company’s own outline of the model and its training campaign.
Three Chinese Models Define the Comparison Set
Reflection’s comparison language assigns different roles to the three Chinese models it names. It says Beam is competitive with GLM-5.2 and close to Qwen 3.8-Max on coding and agent tasks. On raw capability, it says Kimi K3 leads. The performance and efficiency figures in the launch material are Reflection’s own assessment, and the announcement keeps the coding-and-agent comparison separate from the claim about raw capability.
GLM-5.2 also serves as the reference point for Beam’s detailed compute claim. Z.ai’s GLM-5.2 release says the model has a one-million-token context window and is offered under an MIT license. Its repository lists 744 billion total parameters and 40 billion active parameters. Reflection’s announcement names GLM-5.2 as a reference point, while the source disclosures provide the stated configuration figures relevant here.
Placed beside each other, the published parameter descriptions are straightforward: Reflection lists Beam at 501 billion total parameters and 23 billion active parameters; the GLM repository lists GLM-5.2 at 744 billion total and 40 billion active. Z.ai supplies the separate context-window and licensing details. Those numbers describe stated configurations, while Reflection’s comparative language concerns the workload categories in its own announcement.
Qwen 3.8-Max is the other model in Reflection’s coding and agent-task comparison. In Reflection’s Oct. 5 announcement, Qwen 3.8-Max has a narrowly defined role: a model Beam is said to be close to on those tasks. Kimi K3 completes the group with the company’s separate statement that it leads on raw capability. The distinction prevents the three names from becoming a single, undifferentiated ranking.
The Boundary Around the 3–4x Compute Claim
Reflection says Beam used three to four times less inference compute than GLM-5.2 in its reasoning comparisons. The company calculates the figure as estimated forward-pass FLOPs. It is a Beam-versus-GLM-5.2 comparison, not a measure offered for Qwen 3.8-Max or Kimi K3, and it belongs to the reasoning comparisons rather than to every workload named in the launch.
The figure is narrow by design. Reflection excludes prompt prefill, context attention, and serving overhead from its forward-pass-FLOPs estimate. In practical terms, the company is describing a defined slice of inference work, not an all-in serving-cost calculation. That scope is important when reading the 3–4x number beside the model’s broader coding, reasoning, and agentic positioning.
Beam’s public profile is therefore still taking shape around Reflection’s Oct. 5 account: a 501-billion-parameter model with 23 billion active parameters, a disclosed training narrative, and comparison claims aimed at a specified group of Chinese models. The next milestone remains the later-October release of the weights, technical report, model card, and developer artifacts. That package will complete the rollout Reflection outlined when it announced Beam.