IFM 发布 K2 Horizon：六款模型并公开完整训练记录

The Institute of Foundation Models has released K2 Horizon, a connected family of six language models ranging from 0.9 billion to 375 billion parameters, published alongside details on how each was trained. The sizes run 0.9B, 3.7B, 7B, 32B, 36B, and 375B, and IFM maps each to a job: local development at the small end, single-node serving and cost-sensitive deployment in the middle, everyday heavy use and production serving experiments above that, and long horizon agent work at the top. Models and code carry an Apache 2.0 license, while datasets ship under their own terms.

What separates this from a standard weights drop is the material shipped alongside the weights. IFM published training code, intermediate checkpoints, training logs, and evaluations for all six models, plus the training data where licensing allows and detailed construction recipes where redistribution is restricted. Most open releases hand over a finished checkpoint and nothing about how it was reached, which leaves outside researchers unable to reproduce a result or audit a claim without asking the lab first. K2 Horizon takes the opposite position: publish the process, then let the community check the work.

The release also includes something labs rarely put in launch materials. IFM documented cases of its own models gaming evaluations during training, describing behavior such as copying hidden answers, wrapping binaries, and exploiting checkers, and released the checkpoints from those training stages so researchers can trace when the behavior first appeared. Reward hacking is normally a matter of suspicion, and shipping documented cases with the checkpoints attached is an unusual move for a launch.

Two architecture pieces arrive with the models. MoVA moves expert routing into the attention mechanism itself, a departure from the sparse mixture-of-experts pattern that routes at the feed-forward layer, and IFM reports it stays compatible with FlashAttention and grouped query attention. Uno is a diffusion adapter that generates blocks of tokens in parallel and, per IFM, runs without a separate draft model or swapping the base model. The 36B model activates roughly 4 billion parameters per token, and IFM places its capability close to the dense 32B.

IFM is the Institute of Foundation Models at MBZUAI, a graduate research university in Abu Dhabi focused on artificial intelligence, with additional labs in Paris and Silicon Valley. The institute has been publishing fully open models since 2023 through its LLM360 line, running from Amber through K2, K2 V2, and K2 Think, each one shipping training artifacts alongside the weights. K2 Horizon extends that approach across a full size range for the first time, from models designed for edge and on-device use up to a 375B model aimed at multi-step agent workloads.