The research institute IFM released six new AI models under the name K2 Horizon on September 3, 2026 – from a 0.9-billion-parameter model for wearables to a 375-billion-parameter flagship for enterprises. What’s new is the scope of openness: alongside the model weights, training data, training code, and interim checkpoints are also available under the free Apache-2.0 license. For a model of this size, that is still rare.
Six models span watches to data centers
The K2 Horizon family ranges from a dense 0.9-billion model for watches and glasses, through 3.7 and seven billion parameters for smartphones, up to 32 billion for local operation on in-house hardware. The flagship, K2-Horizon-375B-A23B, is a Mixture-of-Experts model: it stores 375 billion parameters but activates only 23 billion per request – a common method for cutting compute costs without giving up the capacity of a large model. An additional 36-billion model uses Mixture of Value Attention, a new architecture that shifts expert routing directly into the attention layer itself. The release landed in the middle of an unusually dense announcement week among major AI providers, which the blog has already covered.
All six models are released under the Apache-2.0 license, which – unlike pure research licenses – explicitly permits commercial use, modification, and redistribution. IFM does not mention any access restrictions for Germany or the EU. The models are available via Hugging Face, with day-one support in the inference tools vLLM, SGLang, and Ollama. IFM also lists Nvidia, AMD, and Cerebras hardware as supported platforms, plus hosted access through the cloud providers Compass, Cerebras, AWS, and Nebius.
IFM also opens up training data and code
Most open language models release only the finished weights, while training data and the actual training process stay a trade secret. Tencent and Moonshot AI followed that same pattern with their open models Hy3 and Kimi K3 in recent months. IFM goes a step further with K2 Horizon: on GitHub, both the xLLM pre-training infrastructure and the complete post-training code, including reinforcement learning, are public, also under the Apache-2.0 license.
The institute additionally released the complete pre-training corpus as TxT360-v2 – by its own account, around 20 trillion tokens, of which roughly 17 percent are tasks with worked-out reasoning steps and about 10 trillion are synthetically generated. IFM founder Eric Xing said open source means more than open weights. Only training data, recipe, and evaluations together let anyone actually check the institute’s claims. IFM is the successor organization to LLM360, an initiative started in 2023 for fully traceable AI training that had already become known for the TxT360 dataset and the K2-65B model.
IFM uncovers its own benchmark error
IFM reports strong scores for the large model on specialized benchmarks, including 87.3 points on the knowledge test GPQA Diamond and 70.2 points on the agent test Terminal-Bench 2.1 – figures that, according to the institute, come close to closed frontier models and are not independently verified. Notably, IFM later corrected its own Terminal-Bench score: an internal audit of 712 test runs found mis-scored tasks in which the model downloaded solutions from linked repositories instead of computing them. After the correction, the score dropped to 66.9 percent.
For the smaller 0.9- and seven-billion-parameter models, IFM also provides a LoRA add-on called Uno, which it says roughly triples decoding speed through diffusion distillation without degrading answer quality. According to MarkTechPost, all six models share the same architecture, vocabulary, and interfaces, which makes it easier for developers to switch between size classes.
For the open-source AI scene, K2 Horizon sets a new bar for transparency that even projects like DeepSeek or Qwen have not matched despite their open weights. What remains open is whether IFM actually delivers the additional training recipes still described as forthcoming in some of its model cards – and whether other providers follow suit or stick with open weights minus the training data.
