Release2026-09-03

IFM released K2 Horizon, a fleet of six open-source AI models. The models range from 0.9 billion to 375 billion parameters. IFM published the weights, code, training data and training methods for each one. The smallest is built for watches and glasses, the 3.7B and 7B for phones. The 32B and 36B models target local hosting and on-premise servers. The 375B model is aimed at enterprise workloads. All six share one architecture, vocabulary and deployment tooling, so developers can move between sizes. IFM claims new best-in-class results at the 0.9B, 3.7B and 7B sizes. Two techniques are named: parallel token generation that roughly triples speed, and a new attention architecture. The models are on Hugging Face, vLLM and SGLang under Apache 2.0.

What changed

Most open model releases share weights only, without training data or recipes.

What it unlocks

Running, inspecting and reproducing a full model family from watch-sized to enterprise scale.

  • six models, 0.9B to 375B parameters
  • 375B flagship, 23B active
  • roughly 3x faster generation
  • Apache 2.0 license

What you need to act on it

  • hardware or an inference partner to host the models

Sources

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.