Reflection AI unveils Beam, its first open-weight model

Reflection AI, a two-year-old startup backed by Nvidia ↗ 2.12%, released Beam on Monday, October 5, 2026. Beam is the company's first open-weight AI model. Reflection says it matches leading Chinese open models on advanced reasoning benchmarks while using much less compute. The company describes Beam as a sparse mixture-of-experts system with 501 billion total parameters, of which 23 billion are active at a time. It was built for coding, reasoning and agentic work.unite+3

The launch puts a well-funded U.S. lab into a race that has mostly been led by Chinese developers such as DeepSeek, Alibaba's ↗ 4.68% Qwen team and Z.ai. TechCrunch reported that the announcement confirmed weekend reporting from Axios that a release was close.techcrunch+1

The performance claims

Reflection says Beam scores about the same as Z.ai's GLM-5.2 on advanced reasoning benchmarks while using "3-4x less inference compute." It also says Beam beats today's leading Western open models. The company described Beam as getting close to Qwen 3.8-Max on coding and agentic tasks. It also acknowledged that Moonshot's Kimi K3 is still ahead on raw capability.quantumzeitgeist+2

Reflection published its own benchmark scores: 80.1 on Terminal Bench v2.1, 80.9 on SWE-Bench Verified and 90.5 on GPQA Diamond. Its compute figures are estimates, not measured serving costs. They leave out prompt prefill and serving overhead, so the company called them approximate. TechCrunch noted that none of these claims have been independently verified.unite+1

The company said Beam was pretrained on 23.8 trillion tokens and has a context window of 1 million tokens. Its reinforcement-learning phase ran for four weeks on 10,500 Nvidia GB300 GPUs and produced more than 100 million rollouts. Reflection calls Beam a "workhorse model" for enterprises, the public sector and developers.techcrunch+1

Release plans

Reflection plans to publish Beam's weights under an Apache 2.0 license later in October. A technical report, a model card and tools for running and fine-tuning the model will come with them. Until then, an early version is open to selected users through a waitlist while final red-teaming continues. The company says it will distribute Beam through hyperscalers and neoclouds.techcrunch+2

Bigger ambitions

Bloomberg reported that the company was founded by two former Google DeepMind researchers. Reflection says it is supplying its open-weight models to the Department of Energy's Genesis Mission, which serves the 17 U.S. national laboratories. According to Axios, the startup's long-term goal is an "AI factory" that would let institutions build their own low-cost AI systems on their own data. Axios also reported that Reflection recently signed deals with Nebius ↘ 4.22% and SpaceX to rent Nvidia servers.axios+2

CEO Misha Laskin has said it will take time for the company's models to reach the most advanced capabilities. "They're kind of like rocket ships," he told CNBC.axios