Reflection AI releases Beam, an open-weight coding model

Reflection AI introduced Beam, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, trained on 23.8 trillion tokens drawn from the web and licensed proprietary datasets. The company, which raised $2 billion a year earlier explicitly positioning itself as an open, US-based alternative to Chinese open-weight labs, called Beam the first in a planned series of models and said it was built as a workhorse for enterprise coding and agentic workloads rather than as a general-purpose chat model.

Reflection said Beam was competitive on coding and reasoning benchmarks with larger open-weight models, naming Zhipu’s GLM-5.2 specifically, while using three to four times less inference compute to reach comparable reasoning scores — an efficiency claim rather than a raw capability lead. These figures were the company’s own, from a model still in a red-teaming and early-access evaluation phase at announcement rather than public release.

Full weights and technical documentation, to be released under the Apache 2.0 licence, were due later in October 2026, with early access available by signup in the meantime. The staggered release — announce, evaluate with selected partners, then publish weights — has become a common pattern among labs shipping frontier-scale open models, reflecting both safety review and the practical difficulty of serving a 500-billion-parameter model at launch.

Beam is Reflection AI’s first model release since its 2024 founding and its $2 billion raise, and places it alongside Mistral as one of a small number of Western labs shipping open-weight models at a scale that competes directly with China’s frontier open-weight releases, an area that labs backed largely by American venture capital had, until this release, mostly ceded to DeepSeek, Zhipu and Alibaba’s Qwen team.