I noted IFM’s K2-Horizon release as a contrast to open-weight releases that provide weights alone: it presents openness as potentially including the whole training recipe. A fully documented frontier training run is rare.
IFM published its K2-Horizon fleet on 16 November.
The fleet comprises six models from a 0.9B dense model to a 375B-parameter mixture-of-experts, built for reasoning, coding and agentic work.
Four of them share one 22-trillion-token training corpus.
Alongside the weights, IFM released the training data, the intermediate checkpoints from every stage, the training logs, and the code.
The released code includes the pipelines that produced around 10 trillion synthetic training tokens.
MoVA attention and Uno adapters speed up decoding by up to 2.2 times without a separate draft model.
The benchmark claims are IFM's own, and few labs could reproduce a 375B run.
I noted the open question: Does open mean the weights, or the whole recipe?