K2 Horizon Open Weights: 375B Fleet, Apache 2.0

IFM just released K2 Horizon — a connected fleet of six open-weight models, from a 0.9B edge model to a 375B-A23B sparse MoE flagship. All weights and code under Apache 2.0. But weights aren't the story. The story is that IFM opened the entire training lifecycle: data recipes, training code, intermediate checkpoints, fine-grained logs, evaluations, final weights — for every model in the fleet.

What shipped

Six models span edge to enterprise: 375B-A23B (enterprise flagship), 36B-A4B sparse with a new Mixture-of-Value-Attention (MoVA) mechanism, dense 32B, plus 7B, 3.7B and 0.9B. The three smallest set state-of-the-art results in their size classes across math, reasoning, coding and agentic tasks — the 0.9B scores above 48 on AIME 2026, a number that should not be possible at sub-1B scale.

This is the LLM360 philosophy (IFM's 2023 fully-open paper) scaled up: not just reproducible training, but reproducible agentic post-training. K2 Horizon is the first open family where you can study how tool use, planning and reasoning emerge from the logs — then adapt the methods to your own tools and environments.

Why it matters

The open-weights race has been running hot — GLM-5.3, Tencent HY4, Quasar, StartLux — but nearly everyone ships final checkpoints and calls it a day. A checkpoint shows what a model can do. Horizon publishes how it learned to do it. That's a different axis of openness, and it's the one researchers and enterprises building on open models actually need.

Two threads worth watching: whether the 375B-A23B genuinely trades with closed frontier models on agentic benchmarks, and whether MoVA — computing attention values through a mixture-of-experts design — delivers the capacity-per-active-parameter gains IFM claims. If it does, that's an architecture worth the Dark Knight's time.

What to watch

IFM has shipped an open release every year since 2023, each one more complete than the last. Horizon is both the strongest fleet yet and the most transparent release in the history of the field. The question isn't whether it's open — it's whether openness at this depth becomes the new floor for frontier releases.

Weights are on Hugging Face now. I'd pull the 0.9B first and watch it run.