Reflection AI has officially unveiled Beam, its first frontier open-weight AI model — a text-only reasoning and coding model the Brooklyn startup says matches China’s leading open models at a fraction of the inference cost. Here’s what the launch means for the open-weight race.
Brooklyn, N.Y. — Monday, October 5, 2026.
What happened
Reflection AI, the two-year-old startup founded by two former Google DeepMind researchers, announced Beam on Monday, confirming weekend reports that a launch was imminent. The company calls it a frontier-class open-weight model and a “workhorse” for reasoning, coding, and agentic tasks — built for enterprises, the public sector, and developers.
The details
Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active, pre-trained on 23.8 trillion tokens with a 1-million-token context window. Reflection says it scores on par with Z.ai’s much larger GLM-5.2 (744B total params) on advanced reasoning benchmarks while using “3-4x less inference compute,” and beats Mira Murati’s Inkling on four reported coding tests (Inkling is multimodal; Beam is text-only).
The caveat that matters most: Reflection’s performance claims have not been independently verified. Weights and full technical details land this month via hyperscalers and neoclouds with open-source library integrations.
The war chest behind it: roughly $4.7 billion raised from Nvidia, Sequoia, and Lightspeed (PitchBook), a $25 billion valuation, and $7 billion-plus in compute deals with SpaceX and Nebius securing Nvidia’s GB300 chips through 2029.
Why it matters
China currently leads the high-end open-weight market. Reflection’s pitch goes further than weights: “AI factories” that let enterprises and sovereign nations build customized, local AI systems on their own proprietary data, backed by Nvidia CEO Jensen Huang — already being tested with Shinsegae Group in South Korea, with hedge funds and trading firms reportedly lining up.
The Tested Intelligence take: open weights are where buyers gain leverage — audit, self-host, fine-tune, no vendor lock-in. Beam’s real test isn’t launch-day benchmarks; it’s whether this month’s weights hold up under independent testing. If the 3-4x efficiency claim survives, enterprise AI teams get a serious new self-hosted option for frontier-class reasoning.
FAQ
When can I download Beam?
Reflection says the weights and full technical details will be released in October 2026, distributed through hyperscalers and neoclouds with open-source library integrations.
How does Beam compare to China’s open models?
Reflection claims Beam matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. Those claims haven’t been independently verified.
What is Reflection’s “AI factory” idea?
A product that would let institutions build their own customized, local AI systems by training Reflection’s models on their own proprietary data — already being tested with Shinsegae Group in South Korea.
Sources: TechCrunch, Axios, Reflection AI.