flow-1 flow-1 is a specialized model trained with reinforcement learning to find errors in AI agent traces. It matches GPT-6-sol-level analysis quality at 23x lower cost, making it possible to monitor every agent run without sampling.
Visit productFeatured media
Loading media…
View on XCosigns Received
“I waited years for this and this has finally happened, Startups are now training their own models for custom use cases that turn out to be much cheaper and faster than frontier intelligence. Intuitively, it makes sense to not use a 400Billion parameter model when all you want to do is solve one specific use case? Its surprising that people haven't done it. I am glad that @skull8888888888 singlehandedly post trained a model to make it 23 times cheaper than Gpt-6-sol at a much much faster inference. My bet is that the a very small base model along with curated post training data from 1000s of runs made using a frontier model is enough to post train your models for this use case. You can then use another frontier model and some skills to optimize your inference pipeline(Read Prime's blog about GLM inference optimization), we have official entered the new world of custom models.”
“Having worked a lot with analytics on top of agent traces, this one is huge if true. Need to play around with this immediately”
