CMF — open weights
· Source: original
Choosing a solution instead of generating tokens — the open CMF model 🎯
We've published CMF — a model with open weights, an analog of Jev. The fundamental difference: it doesn't generate tokens and text, but instead selects a solution class. Plus it can decline uncertain answers instead of producing just anything.
CMF passes unfamiliar cases to an oracle-teacher — a third-party model — and fine-tunes on them. Repeat requests then run without the oracle, independently. This is that very pattern where the teacher is needed only at the start, and after that the model carries on by itself.
Launch — via CLI and API. Adding your own skills is supported. The full pipeline has been ported to Apple Metal and Vulkan.
⚡️ Measurements were taken on an RTX PRO 4000 Blackwell: median latency — 1.10–1.24 ms in a continuous series. The article includes a comparison of CMF with Jev and Laya.
The material is dated 28 September 2026. The full breakdown and measurements — at https://habr.com/ru/articles/1087552/
🤖 Interested in AI agents and automation?
Prompts for building AI agents and automations — read up on the topic:
🔗 The entire prompt library · "AI Agents" category
A ready-made product on the topic: Prompts for Programmers — grab it and apply it right away.