K2 Horizon

What it is

K2 Horizon is a family of six language models released by the Institute of Foundation Models (IFM) on 3 September 2026 ^[1]. The fleet spans from 0.9 billion to 375 billion parameters, so it covers models small enough to run on personal devices as well as a very large one ^[1]. The release is unusually complete for a model launch: it ships weights, code, training data and training recipes, and both the models and the code are published under the Apache 2.0 licence, which allows commercial reuse without a separate agreement ^[1]. IFM calls it the largest fully open-source model launch in AI history; that is the publisher's own description rather than a figure checked by an outside party ^[1].

One architecture, six sizes

All six models share a single architecture, vocabulary, training method and deployment tooling. The one exception is the 0.9B model, which uses a smaller vocabulary ^[1]. Keeping one design across the range means a change made for one model carries over to the others, and a team that learns the deployment tooling once can use it for every size in the family.

Two of the models are built so that only part of the network is used for each token. The 375B model activates 23 billion parameters per token, and the 36B model activates 4 billion ^[1]. In designs of this kind the compute spent on each token stays far below what the total parameter count would suggest, because most of the weights sit idle for any given piece of text. The 36B model goes further with a mixture-of-value-attention design that IFM says improves reasoning without extra computation ^[1].

The family also includes a technique the release describes as diffusion distillation, which generates blocks of tokens in parallel rather than strictly one after another; IFM claims it runs about three times faster ^[1]. Most language models write output sequentially, one token at a time, so producing a whole block in one pass is the main source of the speed-up claimed here ^[1].

Where each model is meant to run

The smallest and the mid-sized models are aimed at consumer hardware. The 0.9B model targets watches and glasses, while the 7B model is meant for phones ^[1]. That pairing sets out the intended spread of the fleet: a very small model for devices with tight memory and power budgets, a phone-sized model for everyday assistant work, and larger models for servers.

Serving is available through vLLM and SGLang, two widely used open-source inference frameworks, with hosted access from Cerebras, Nebius and Compass ^[1]. Providing both self-hosted and hosted routes means a team can start on someone else's infrastructure and move to its own machines later without changing the model itself.

Place in the line

Open releases are usually judged on how much of the pipeline is published, not only on benchmark scores. K2 Horizon publishes data and recipes alongside weights and code, which is what makes the models reproducible by outside groups rather than merely downloadable ^[1]. The Apache 2.0 licence removes the usage restrictions that often accompany open weights ^[1].

The fleet was designed to be adapted. Jais 3, released the following month, was the first adaptation of the family to another language ^[2], which shows the intended pattern: take a shared base, then tune it for a language or domain. Because every model in the family uses the same architecture and tooling, such adaptations can start from whichever size fits the target device ^[1].