This Week in AI: Open Weights Kept Quietly Winning

Liquid AI released d1-3B and d1-omni-600M on October 7 as open weights for local deployment. The 3-billion-parameter model takes text and images and returns a structured decision in a single pass instead of generating prose, which is exactly what you want for routing and classification without a cloud round-trip.

Mistral opened a public preview of Large 4 on October 6, a 1.05-trillion-parameter mixture of experts with 49 billion active parameters and a one-million-token context. The open-weights release is scheduled for late October under Apache-style terms, which matters because it lets companies self-host a frontier-class model.

Reflection AI's Beam and a batch of smaller decision models from Cloudflare, AWS and Perplexity all landed in early October, pulling the 'decide, don't generate' pattern into the mainstream. These are cheap, fast, and increasingly good enough to sit in front of the big models and decide what actually needs their attention.

The throughline is that 'local and open' stopped meaning 'worse.' For the daily, repetitive, privacy-sensitive calls, a model on your own machine is now a real default, not a hobbyist's compromise.

Microsoft's MAI-Code-1.1 Flash, a 137-billion-parameter coding model built for on-device execution, fits the same trend: the laptop is becoming a legitimate place to run serious AI, and that changes who your data ever leaves.