K2 Horizon ships six fully open models from 0.9B to 375B

The Institute of Foundation Models, part of MBZUAI, released K2 Horizon, a set of six models sized 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B. What sets it apart from most "open" releases is how much is actually open: weights, training code, training data, benchmarks, and deployment recipes, all under Apache 2.0, with day-zero support in vLLM, SGLang, and Ollama.

The strongest claims are at the small end. The 0.9B, 3.7B, and 7B models set new state of the art at their respective scales, with the 0.9B model scoring above 48 on AIME 2026 while keeping usable reasoning, tool-use, and agentic behavior. The team credits a Mixture-of-Experts design that computes attention efficiently, which is why the 36B-A4B model, with only 4B active parameters, performs above what its active count would suggest. The 32B and 375B-A23B models are pitched as competitive within their own size classes rather than as outright frontier leaders.

A connected fleet from sub-1B to 375B, all trained with the same published recipe, is the useful part here. It gives people a single, reproducible line to study rather than one isolated checkpoint.

Why it matters

If you run models on your own hardware, a fully open small model that scores 48 on AIME is worth a real look for edge and on-device work, where you cannot ship a frontier API. Because the training code and data are published, you can also fine-tune or audit it rather than treat the weights as a black box.