open models are free to download and you can take them anywhere. so the meter moved to the things you can’t take with you when you leave: the cache, your memory, and the gateway. when one company owns the router and the payment rail, it knows exactly how much it would cost you to leave. run a router you control. route on cost. keep a second model warm. stateofopensource.ai

thank you to everyone who joined me and @mkrigsman on @cxotalk last week. the conversation was insighful. the questions were sharp. watch the replay here: cxotalk.com/episode/mozilla-…

capability gains in AI no longer require a new model. the model and the harness still matter, but real results happen when working on an open base. in july, DeepSeek ran a post-training pass that lifted V4-Flash 21 points on Terminal Bench and 10 on the  AA index. same model. same architecture. same size. stateofopensource.ai/

an agent can prove who it is. no one can say what it’s allowed to do. MCP went from ~2M to 110M+ monthly SDK downloads in 21 months. meanwhile ~21% of companies say their agent governance is mature. stateofopensource.ai/

serious AI used to live on someone else’s data center. you rented it by the token. that’s changing. Capable open models now run at every hardware tier, from a gaming PC to a small data center. top models improved by adding hardware. smaller models improved on the same hardware. capability exists at every tier. the complete overview is in the State of Open Source AI report: stateofopensource.ai/

the best open model is only three behind closed. open models went from something you could rent, to something you could run. now, you can improve them too. full data in @mozilla's state of open source AI report. stateofopensource.ai/

this spring, the best closed model scored 60 and the best open models 54. Aayear earlier, the leading open model scored 22. the frontier still leads on the hardest problems — and for the work most builders actually ship, where price, control, and deployability decide, the data does not say “promising. read the full piece from @tech_eu tech.eu/2026/09/18/open-ai-c…

on august 10, mark zuckerberg published 6,500 words arguing that AI should belong to everyone. that same day, meta shipped an open-weight model. i downloaded it before i finished reading the essay. so can you. my @guardian op-ed argues that before you sign the lease on AI that belongs to someone else, ask 3 questions. can i be evicted? what can i take with me when i leave? who keeps the place standing when the landlord loses interest? read the full piece 👇 theguardian.com/commentisfre…

excited to announce that I'll be sitting down with @jonigl for the very first owners not renters community live stream. join us on August 11am, ET. come prepared for a lively discussion about open source ai. This event is for ONR subscribers only. free to join — just make sure you're subscribed to the owners not renters newsletter. keep an eye on your inbox for more information.

june 12: washington barred any foreign nationals from using anthropic's frontier AI model. anthropic couldn't verify user nationality in real time, so the model went dark. for everyone. my @WSJopinion op-ed: ai shouldn't require a permission slip. the defining question for the future of AI isn't leaderboard usage. it's whether june 12 happens to you next. wsj.com/opinion/ai-shouldnt-…

developers are choosing open source AI. they just can't ship it. 79% use open models. only 51% get them into production. 🧵