@adlrocha - From RLHF tro RLCD

How can decision models be so fast compared to LLMs? OpenAI has said its Decisions API runs on a specialised GPT-6 Luna, and all TypeSafe has said about Jev is that it has a “new architecture” and a “parallel sampler”. So does this mean that these models are a completely new architectures, or that they are built leveraging LLMs? Thankfully, Cloudflare’s Clef, released two weeks after Jev, and by being an open weight model, it comes with the recipe and the underlying architecture that has allowed me (and the whole community) to understand how this models come to be.A month ago, writing about DeepSeek-V4.1-Flash, I described one of its trick as splitting the model in two: “first half reads, second half writes”. Reading is cheap and writing is expensive, so this new DeepSeek model avoided paying the writing price where it wasn’t needed. A decision model is based on this same idea by completely throwing the writing half away. Over the last two posts I covered what decision models are for…