IFM releases K2 Horizon open models, weights up to 375B

The Institute of Foundation Models has released K2 Horizon, a six-size family of open-weight models under Apache 2.0, from a 0.9B dense model up to a 375B-parameter mixture-of-experts with a 512K context window. Weights are live on Hugging Face now; the training data and code are promised, not yet public.

Key facts

• The flagship K2-Horizon-375B-A23B stores 375B parameters and activates 23B per token, with a native 524,288-token context window from midtraining onward.
• The family ships in six sizes on Hugging Face: 0.9B, 3.7B, 7B, 32B, a 36B-A4B mixture-of-experts variant, and the 375B flagship, with GGUF and FP8 builds for most.
• All models carry the Apache 2.0 license, the most permissive standard option for commercial and derivative use.
• IFM's own benchmark table puts the 375B model at a 1,441 Elo on GDPVal-AA real-world professional tasks, behind closed frontier models but ahead of every open model listed, including the 550B Nemotron 3 Ultra.
• The model card names IFM/K2-Horizon-Pretrain-Data and IFM/K2-Horizon-Midtrain-Data as training datasets, but neither dataset repo was public at release time.

What happened

The Institute of Foundation Models, the Abu Dhabi-based lab behind the K2 and Amber model lines, published the K2 Horizon family on Hugging Face. The release covers six models: dense versions at 0.9B, 3.7B, 7B, and 32B parameters, a 36B mixture-of-experts variant activating 4B per token, and the flagship 375B-A23B mixture-of-experts. Every model shares the same 524,288-token context window, which IFM says is native from the midtraining stages rather than bolted on afterward.

The licensing is as open as this class of release gets. All K2 Horizon models are Apache 2.0, and the card commits to releasing training data, recipes, training code, and intermediate checkpoints. That last item is the unusual one. Publishing intermediate checkpoints lets outside researchers study how capabilities emerge across training, not just probe a single finished artifact.

The practical serving story is already in place. IFM published GGUF builds of the four dense sizes for llama.cpp-style local inference, and FP8 quantizations of the larger models. The card lists day-zero support for vLLM, SGLang, and Ollama, so the models slot into the standard open-serving stack without custom patches.

Why it matters

Two things set this release apart from the weekly drumbeat of open weights. The first is the stated commitment to full-lifecycle openness: weights, data, recipes, code, and checkpoints. If IFM delivers, K2 Horizon would join a very short list of frontier-adjacent models that outside teams can genuinely reproduce and audit rather than just download. As DeAI has covered in the difference between open-weight and open-source models, most "open" releases stop at weights and never publish the data or the recipe.

The second is the size spread with a single context window. A 0.9B model and a 375B mixture-of-experts that both handle 512K tokens give builders a consistent interface across five orders of magnitude of compute. That matters for the tiered-inference pattern this publication tracks in decentralized AI inference: small models at the edge, flagship models in a datacenter, one tokenizer and one context contract.

A necessary caveat: every performance figure in this story comes from IFM's own model card. The 1,441 GDPVal-AA Elo, the agentic benchmark leads over larger open models, the 512K context claim — none of it has been independently verified. The lab's claim that the 375B model beats open models up to 2.6 times its size is exactly the kind of provider self-report DeAI labels as a claim until reproducible evals land.

Background

The Institute of Foundation Models is the Abu Dhabi lab behind the earlier K2, K2-Think, Amber, and Crystal models, and it has a track record of shipping data as well as weights: its MegaMath and TxT360 datasets are public and widely used. That history lends some credibility to the promise that K2 Horizon's training data will follow, but the dataset repos named in the card, IFM/K2-Horizon-Pretrain-Data and IFM/K2-Horizon-Midtrain-Data, returned no public listing at release time. The weights are open; the full pipeline is not yet.

K2 Horizon lands in the middle of the busiest stretch of open-weight releases this year. It follows Alibaba's Qwen 3.8 pairing of a frontier API with an open 27B model, and OpenBMB's fully reproducible MiniCPM5-2B, which DeAI covered for shipping weights, data, and recipes together. The pattern is a slow industry shift from "open weights" as a marketing term toward reproducibility as the actual bar, a theme the World Economic Forum also picked up this week in framing open-weight AI as owned infrastructure rather than rented access.

What's next

Watch the IFM Hugging Face organization for the pretrain and midtrain dataset repos and the training code; their appearance converts this from a weights release into a reproducible one. Independent benchmark reproduction, especially of the GDPVal-AA Elo and the agentic tool-use claims, is the real test of the 375B flagship's position. And the promised intermediate checkpoints will show whether capability emergence in a 375B mixture-of-experts looks anything like the dense models researchers have studied to date.