# K2 Horizon opens more than model weights — including training data, code and intermediate checkpoints The Institute of Foundation Models has released a six-model K2 Horizon family spanning 0.9B to 375B parameters with Apache-2.0 model code and an unusually broad set of training and evaluation artifacts. Some model cards still mark parts of that release pipeline as forthcoming, so 'fully open' is not yet uniform across every asset. K2 Horizon is notable less for another benchmark claim than for reproducibility: IFM is publishing model weights, architecture, training code, data or construction recipes, evaluation resources and intermediate training material instead of stopping at a final checkpoint. - Status: Active - Published: 2026-09-04T10:15:15+12:00 - Updated: 2026-09-04T10:15:15+12:00 - Categories: Artificial Intelligence, AI Models, Open Models, Inference & APIs - Tags: Apache 2.0, Institute of Foundation Models, K2 Horizon, open models, reproducibility, training data - Canonical HTML: https://beyondthe.news/dossiers/k2-horizon-open-models-training-data-code-checkpoints ## What changed The Institute of Foundation Models (IFM) released K2 Horizon on September 3 as a family of six models ranging from a 0.9B dense model to a 375B mixture-of-experts flagship. IFM says the release opens the training lifecycle rather than only final weights: architecture details, model code, training configurations, data or detailed data-construction recipes, evaluation resources, logs and intermediate checkpoints are part of the published/release plan. The model and code license is Apache 2.0, while datasets retain their own applicable licenses. Hugging Face already hosts the model collection and public K2 training-data resources, but individual model cards still say some intermediate checkpoints, training data or training code 'will be released', so the completeness of the open package varies by model at launch. ## Why it matters Open-weight releases let builders run a model; reproducible releases let researchers inspect how the model got there. K2 Horizon gives developers and researchers substantially more material for auditing training choices, studying capability changes across checkpoints and building derivative models without reverse-engineering an opaque recipe. The smaller 0.9B, 3.7B and 7B variants also make that research possible below data-center scale. The practical caveat is that IFM’s performance claims are self-reported and some promised artifacts are still propagating across the public repositories, so builders should verify the exact files available for the specific K2 model they plan to use. ## The family spans edge-sized to large sparse models K2 Horizon includes 0.9B, 3.7B, 7B and 32B dense models plus a 36B model activating roughly 4B parameters and a 375B flagship activating roughly 23B. IFM positions the smaller models for constrained and on-device use and the sparse models for larger local or enterprise inference. ## The open release targets the training lifecycle, not just inference IFM says it is publishing training data or detailed construction recipes, code, configurations, mixture information, fine-grained logs, evaluation resources and intermediate checkpoints. This makes it possible to study how capabilities change during training instead of evaluating only a final weight file. ## The public repositories show a release still being filled in The Hugging Face collection is live and includes final model checkpoints plus K2 training-data resources. At the same time, several individual model cards say intermediate checkpoints and some training assets will be released. Builders should therefore treat 'fully open' as the intended release architecture and inspect the current repository state before assuming every promised artifact is already downloadable. ## Benchmark claims still need independent replication IFM reports strong reasoning, coding, agentic and long-context performance across the family, including state-of-the-art claims at smaller sizes. Those comparisons are project-authored. The unusually open training package makes independent replication easier, but it does not replace it. ## Key details - K2 Horizon launched September 3, 2026. - The family contains six models spanning roughly 0.9B to 375B parameters. - The 36B MoVA model activates about 4B parameters per token; the 375B model activates about 23B. - Several models advertise context windows up to roughly 512K tokens. - IFM says model weights, code, training data or detailed construction recipes, methodologies, evaluation resources and intermediate checkpoints are part of the open release. - Models and code are released under Apache 2.0; datasets use their applicable data licenses. - The K2 Horizon collection is live on Hugging Face. - Some individual model cards still describe intermediate checkpoints, training data or training code as forthcoming, so asset availability is not uniform at launch. - Reuters independently confirms the release and its emphasis on training-data/code transparency. ## Builder takeaways - Choose the exact K2 variant based on memory/compute constraints rather than treating the family as one deployment target. - Pin model and dataset revisions when reproducibility matters because the launch repositories are still receiving artifacts. - Audit the license of each dataset separately; Apache 2.0 on model/code does not automatically apply to all training data. - Use the intermediate-checkpoint and data releases for research/auditing only after verifying that the promised assets for your chosen model are actually present. - Reproduce IFM’s benchmark claims on your own harness before substituting K2 into production model routing. ## What to watch - Completion of the promised intermediate-checkpoint and training-code/data uploads across every K2 variant. - Independent benchmark replication, especially for the smaller models and the 36B-A4B architecture. - Inference-provider support and real-world cost/latency measurements for the sparse models. - Derivative fine-tunes or research that use K2’s intermediate checkpoints to study training dynamics. - Whether other open-model producers respond with similarly complete training artifacts. ## Uncertainties - IFM’s launch language describes the family as fully open, but some individual Hugging Face model cards still mark certain training assets and intermediate checkpoints as forthcoming. - Performance claims are primarily produced by IFM and have not yet had time for broad independent replication. - Dataset licensing and redistribution constraints differ across the training corpus even when construction details are disclosed. - The largest models still require substantial hardware despite their sparse activation. ## Sources - [Introducing K2 Horizon: Frontier Performance, Radically Open](https://ifm.ai/blog/k2/) — Institute of Foundation Models · primary · 2026-09-03T00:00:00+12:00. Primary technical launch describing the six-model family, open training lifecycle, architecture and licenses. - [K2 Horizon Press Release](https://ifm.ai/k2/press-release/) — Institute of Foundation Models · primary · 2026-09-03T00:00:00+12:00. Primary launch summary for model sizes, deployment targets, availability and licensing. - [K2 Horizon collection](https://huggingface.co/collections/IFM/k2-horizon) — IFM / Hugging Face · primary_repository. Live model/data collection showing the currently published K2 assets. - [K2-Horizon-7B model card](https://huggingface.co/IFM/K2-Horizon-7B) — IFM / Hugging Face · primary_model_card. Primary model documentation illustrating context, training-artifact commitments and reproducibility guidance. - [Abu Dhabi AI institute releases fully open-source models with training data, code](https://www.reuters.com/world/middle-east/abu-dhabi-ai-institute-releases-fully-open-source-models-with-training-data-code-2026-09-03/) — Reuters · independent_reporting · 2026-09-03T00:00:00+12:00. Independent confirmation of the six-model release and the unusual breadth of disclosed training material.