Open Weights Had Its Best Month
Beam, Mistral Large 4 and Qwen drops put front-tier models in operator hands.
October was unusually good for people who want to run models themselves. Reflection AI's Beam (501B sparse MoE, 1M context) announced open weights under Apache 2.0. Mistral Large 4 -- a trillion-parameter European multimodal model -- promised downloadable weights by month's end. Alibaba shipped Qwen-Image-2.1-Turbo and open-sourced the QwenGyre agent framework; Microsoft built its Decision-1 on open-weight Qwen3.5-9B.
Why it matters: open weights mean the research community, not the vendor, audits the claims. A half-trillion-parameter model in the wild gets stress-tested within weeks, publicly.
The honest catch is hardware. These models don't run on a laptop. But the weights existing flips the default from 'trust us' to 'check it.'
For regulated and sovereignty-minded buyers, this is the month the realistic self-host option arrived. Expect more shops to weigh capability against where the servers live.