Between November 15 and 17, 2026, IFM published K2-Horizon, a fleet of six open language models that runs from a 0.9B dense checkpoint to a 375B mixture-of-experts, and filed the long-form technical report as arXiv 2611.21003 on November 16. The pitch is not the size ladder alone. IFM says it releases the pipelines behind the models too: the training data, intermediate checkpoints, training logs, evaluation code and infrastructure code, at ifm.ai/k2 and on Hugging Face and GitHub, in the LLM360 style of openness. Four of the models were trained on a shared 22 trillion-token corpus, and IFM says the pipelines that produced roughly 10 trillion synthetic pretraining tokens ship with their documentation, so the data recipe can be inspected and criticized rather than taken on trust.
The fleet, and the two ideas that matter
The six sizes are a 0.9B dense model for edge and tool-use prototypes, 3.7B and 7.0B dense models for mid-size reasoning and scientific coding, a 32B dense model as the largest dense checkpoint in the family, a 36B-A4B mixture-of-experts, and the 375B-A23B flagship for coding and agentic work. Two modeling ideas carry the release. The first is Mixture-of-Value Attention, or MoVA, which applies mixture-of-experts logic to attention value projections instead of only to feed-forward layers. IFM claims the 36B-A4B MoVA model nearly matches a 32B dense model trained on the same data while activating about one seventh of the parameters, a claim to validate on your own agent harness rather than in a leaderboard screenshot. The second is Uno, a lossless diffusion adapter that speeds decoding up to 2.2 times without a separate speculative draft model. The training stacks ship as xLLM for dense and MoE training and RL360 for post-training, and IFM lists vLLM, SGLang and Ollama as deployment paths. IFM states the models lead or stay competitive with strong open models at each size class on agentic and coding benchmarks, but until independent labs publish harness logs, treat those cells as vendor-reported.
How to try it without a GPU cluster
Start with the 7B or 32B size if you do not run a cluster, and use the 0.9B model only for edge prototypes IFM documents. The 375B-A23B flagship is a multi-GPU or hosted inference job, not a laptop download, so treat that headline as infrastructure news unless you already serve comparable open MoE models. Read the data card and the license files bundled with each checkpoint before any commercial fine-tuning, and mirror only the size class you need, because publishing terabytes of data and checkpoints helps researchers but burdens mirrors. Two checks that are not about model quality: IFM is Abu Dhabi-based, so some teams need to map data residency and export rules before mirroring full corpora, and the name collided with rumors before the primary docs existed, so anchor on ifm.ai/k2 and the 2611.21003 paper, not repost threads. Our Inkling 2 Small explainer covers the other big open-weight drop this month and its VRAM math, our vLLM serving guide covers the engine IFM lists for running these models, and our local LLM guide covers the single-machine end of the same problem. If you want the GPU sizing, the serving layer, and the monitoring around an AI workload checked by someone who runs infrastructure for a living, book a free infra audit with DevOpsBY.
Disclaimer: This post is educational only, not financial advice. Figures are as reported by the sources on 2026-11-17 and may change - always confirm version numbers, requirements, and prices in the official docs before acting on them.
