Reflection AI Announces Beam, a 501B Open Weight Model Due Under Apache 2.0 in October

Reflection AI on October 5, 2026 announced Beam, its first open weight model, a sparse mixture of experts model with 501 billion total parameters and 23 billion active per token. The weights, a technical report and a model card will follow later in October under the Apache 2.0 license, the company said.

Reflection, a Brooklyn based startup founded in 2024 by former Google DeepMind researchers, has positioned itself as a Western answer to the open weight models from Chinese labs such as DeepSeek and Qwen, The Decoder reports. Reflection publishes model weights but keeps its training data and pipelines proprietary (The Decoder). TechCrunch notes that Axios had reported over the preceding weekend that a launch was close, and Mistral AI made a similar weights promise for Mistral Large 4 a day later.

What did Reflection AI announce?

Reflection AI announced Beam on October 5, 2026 as a text only model built for coding, reasoning and agent workloads, and said it is still in final red teaming. A mixture of experts model splits its parameters into many "experts" and runs only a few for each token, so it computes like a much smaller model.

The company says Beam supports a context window of 1 million tokens and has a reasoning effort parameter: lower settings give shorter answers, higher settings let the model think longer on hard tasks. In its announcement, Reflection says Beam is "the first model in a series" and that it is already training a successor.

The company says that during reinforcement learning on reasoning, software and terminal tasks, Beam got better at web browsing even though no browsing tasks were in that training mix.

How does Beam compare with Chinese open models?

Reflection says Beam scores on par with Z.ai's GLM 5.2 on advanced reasoning benchmarks while using three to four times less inference compute. It also says Beam is "approaching" Qwen 3.8 Max on coding and agent tasks, and that models such as Kimi K3 "remain ahead on raw capability."

The efficiency claim rests on Reflection's own estimate. The company calculates compute as roughly twice the active parameters times the average number of generated tokens, using data from Artificial Analysis and DataCurve, and calls the result "an approximate compute comparison rather than measured inference cost." It also measures against GLM 5.2, although its own benchmark table lists the newer GLM 5.3. For scale, TechCrunch puts GLM 5.2 at roughly 744 billion total and 40 billion active parameters.

Counted over the full table, Beam trails Qwen 3.8 Max on all 13 rows where both have a score and Kimi K3 on 13 of 14; the exception is tau3 banking, where Beam scores 38.0 to Kimi K3's 37.1. GLM 5.3 scores higher than Beam on all 12 rows where both are listed, and DeepSeek V4.1 Flash on 7 of 9. Against the older GLM 5.2, Beam leads on 8 of 13 shared rows. Against Inkling, the open model that Thinking Machines Lab released in July, Beam scores higher on the four coding tests where both report results, according to TechCrunch, which adds that Inkling is multimodal and Beam is text only.

TechCrunch stresses that none of these claims has been independently verified. Artificial Analysis, an independent benchmarking firm that Reflection gave early access, said in an October 5 post quoted by Implicator.ai: "Early indicators suggest Beam will be one of the most token-efficient open models we've seen for its level of intelligence." No full independent evaluation has been published.

How was Beam trained?

Reflection says it pretrained Beam on 23.8 trillion tokens from the web, public sources and licensed proprietary data, in under four weeks on 6,144 Nvidia GB300 GPUs. It then ran reinforcement learning, a training phase where the model attempts tasks and is rewarded for solving them, on 10,500 GB300 GPUs for four weeks.

That phase produced more than 100 million rollouts, or task attempts, drawn from nearly one million coding, agent and science environments, according to the company. Reflection calls it "one of the largest scale RL runs conducted by any open lab to date" and says scores were still rising when it ended.

For safety, Reflection says it trained a separate model on its safety and behavior rules and merged it into Beam through distillation, a technique where one model learns from the outputs of others. It plans to publish safety results in the technical report and to open source the safety evaluations it built.

Compute is central to the company's strategy. TechCrunch reports that Reflection signed deals worth more than $7 billion with SpaceX and Nebius this summer for access to GB300 chips through 2029.

Who is Reflection selling Beam to?

Reflection is aiming Beam at enterprises and governments through what it calls "AI factories," local systems trained on an institution's own data, and is testing that idea with Shinsegae Group in South Korea, TechCrunch reports. The company says it will launch with distribution partners, integrations in open source libraries, and tools for running, evaluating and fine tuning the model.

The scale contrasts with Aleph Alpha's Kolibri, a 78 billion parameter open model released on October 3, 2026, which targets European public administration and industry.

What we don't know yet

• The exact release date of the weights, technical report and model card.
• How Beam performs in full independent evaluations, including Artificial Analysis's final results.
• What hardware Beam needs to run, and whether quantized versions will be offered.
• Pricing and availability through cloud partners, including in the EU.