AI Pulse · AI Trends Pulse

A new open weights model called Ornith-1.5 just showed up, and the training approach behind it is worth pausing on. Instead of a fixed team of researchers designing every training task by hand, the model proposes its own tasks, builds the tools it needs to solve them, and then feeds its own results back in as new training data. It teaches itself, in other words. There are three sizes, a 9B, a 35B mixture of experts, and a 397B mixture of experts, all released under an MIT license, which means anyone can use, modify, or build products on top of them without asking permission.

Here’s the part that matters for owners. The 397B version reportedly benchmarks close to Opus 4.8, which is one of the stronger closed models out there right now. And the small 9B version, once quantized, is said to run on a phone. If that holds up under independent testing, it changes the math on where AI capability lives. You wouldn’t need a cloud subscription or a big compute bill to get useful model performance for certain tasks. You could run something on a company laptop or a field device instead.

This is still fresh and the benchmark claims haven’t been independently verified yet, so treat the parity claim as a strong signal, not a settled fact. But the self-improving training loop is a real shift in how these models get built, and open, MIT-licensed weights at this capability level lower the barrier for smaller companies to run their own models instead of renting someone else’s. That’s exactly the kind of development we track closely, because knowing which models you can run in-house versus which still need a vendor is the kind of thing we build into an AI command centre, so you’re not guessing when the landscape shifts. For more, see the original report.