Reflection AI launches Beam to target Chinese rivals

What did Reflection AI announce about Beam?

Reflection AI said it launched Beam on October 5, 2026 as its first open-weight model, while also saying the model was still undergoing final red-teaming and evaluations. That means the company’s announcement was an early-access release, not a completed public weights release.

The company described Beam as a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active parameters. Reflection said it is targeting coding, reasoning, and agentic workloads, and that the model has a 1 million-token context length. The company also said Beam will be released under an Apache 2.0 license.

How is Beam supposed to compete?

Reflection AI said Beam was built to reduce token cost and inference-time compute relative to rivals, according to TechCrunch. The company framed the launch as a bid to match leading Chinese models at lower compute cost.

Reflection also said Beam was pretrained on 23.8 trillion tokens from web and proprietary licensed datasets. For reinforcement learning, the company said it used 10.5K Nvidia GB300 GPUs over four weeks and generated more than 100 million rollouts. Those details matter for founders because they indicate both the scale of the training run and the amount of infrastructure required to push the model toward task performance.

Reflection said Beam’s weights, technical report, model card, and developer artifacts are scheduled for release later in October 2026, after the final review process finishes. TechCrunch reported that the company planned distribution through hyperscalers and neoclouds, along with integrations across open-source libraries, when Beam launches.

Who is Reflection AI building Beam for?

Cadena 3, citing the company, reported that Beam is aimed at enterprises, governments, and developers as a customizable AI system. Reflection also said it is positioning Beam and future models for enterprise and sovereign customers through so-called AI factories trained on customers’ own data.

The company said it had already begun testing a sovereign AI factory partnership with Shinsegae Group in South Korea. In that model, the point is not just to sell a general-purpose model, but to package infrastructure and customization around customer-owned data and deployment requirements.

Reflection AI was founded in 2024 by former Google DeepMind researchers, according to Cadena 3. The same report said the company has raised about $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, and that its latest round valued the company at $25 billion pre-money. Cadena 3 also reported that Reflection signed more than $7 billion of agreements with SpaceX and Nebius for access to Nvidia GB300 chips through 2029.

What performance claims did Reflection make?

Reflection said Beam was competitive with Z.ai’s GLM-5.2 and was closing in on Qwen3.8-Max for coding and agentic tasks, according to The National. Those are company-attributed claims rather than independently verified benchmark results in the available reporting.

For AI builders, the important takeaway is that Reflection is pairing an open-weight distribution strategy with a very large training and inference stack. The company is also signaling a go-to-market approach that mixes direct model release with enterprise deployment channels and infrastructure partners.

What happens next is straightforward: Beam’s technical materials and weights are supposed to arrive later in October, if the company completes its stated review process on schedule.