K2 Horizon: 6 fully-open models to 375B, training data included
MBZUAI's IFM ships all six under Apache 2.0 — weights, corpus and recipe — reproduce or fine-tune, not just run. The 375B is a 23B-active MoE at 512K.
Copy markdown
Open weights, and the data to rebuild them
IFM released all six K2 Horizon models under Apache 2.0 — and unlike almost every 'open weights' drop, it ships the pre-training corpus, recipe and training code alongside the checkpoints (some flagship artifacts still rolling out). That means you can audit exactly what went in, reproduce a run, or continue pretraining on your own data, not just call a black box.
The 375B flagship, by the numbers
K2 Horizon 375B is a sparse MoE — 375B total, 23B active per token — with a native 512K context. It posts 70.2% on Terminal-Bench 2.1, 67.7% on MCPMark and 65.3% on Toolathlon, and lands at 47 on Artificial Analysis's Intelligence Index — a 30-point jump over its predecessor, with agentic tool use as its strong suit.
Pick your VRAM: 0.9B to 36B too
The fleet scales down to real hardware: a 7B that IFM calls the best model under 10B, a sparse 36B-A4B tuned for cheap local hosting, plus 32B, 3.7B and 0.9B options. One consistent family lets you prototype on the small model and scale up without re-plumbing your stack.
How to run it this weekend
Weights are on Hugging Face now, with day-one support in vLLM, SGLang and Ollama, plus hosted access via Cerebras, Nebius and Compass. A concrete build: fine-tune the 7B on your own docs for a self-hosted assistant, or put the 375B behind MCP as an agent brain you fully control.