Open weights quietly won October 2026, and most people missed it
Three labs shipped or promised downloadable models in one month. The 'only closed models matter' era cracked.
Everyone watched the flagship benchmarks, so they missed the more useful shift: October 2026 was the month open weights stopped being a hobbyist curiosity. Three labs moved real weight into downloadable form.
Mistral previewed Large 4, a 1-trillion-parameter model it described as open-weight, with weights promised by month's end. Moonshot's Kimi K3.1 Pro scored in the top tier on public indexes and committed to open weights by early November. DeepSeek shipped V4.1-Flash with an MIT license and a million-token context window.
Why this matters more than a leaderboard: a downloadable model means you can run it inside your own compliance boundary, fine-tune it on private data, and stop paying per token for steady workloads. For any team with volume, that math beats a hosted API fast.
The catch is real. Open weights still need GPUs, ops, and someone who can keep a model patched. The cloud APIs did not disappear; they lost their monopoly on 'the good model.' Pick hosted when you want zero ops, pick open when control and cost matter more than convenience.
I would not bet the farm on either camp. But a year ago this menu did not exist. That is the story.
Frequently asked questions:
Q: Is Mistral Large 4 actually open source?
A: Mistral calls it open-weight; the weights were promised in preview, not shipped on day one. Check before you build on it.
Q: Can I run these on my own machine?
A: The smaller ones, yes. The trillion-parameter ones need real hardware or a hosted open-weight provider.
Q: Do open weights beat GPT-6 on quality?
A: Not at the top, but for many tasks the gap is now small enough that cost wins.