Skip to content
Chain AI Wire
AI Models

K2-Horizon: Six Fully Open AI Models, From a 0.9B Edge Model to a 375B Flagship

IFM published the full K2-Horizon technical paper on arXiv on 16 Nov 2026: six fully open models from 0.9B to 375B-A23B, with training data, checkpoints, logs and code.

Six glowing glass spheres rising like a mountain ridge, illustrating the K2-Horizon fleet of six fully open AI models from 0.9B to 375B parameters
K2-Horizon: six fully open AI models from a 0.9B edge model to a 375B flagship, with the full technical paper published on arXiv on 16 November 2026.

Abu Dhabi’s Institute of Foundation Models has published the full technical paper behind K2-Horizon, its fleet of six fully open language models, on arXiv. The paper, arXiv:2611.21003, submitted on 16 November 2026, is the document that turns an open-weights release into something researchers can actually check: training data, intermediate checkpoints, training logs and the code that built the models are all part of the release.

At a glance

  • K2-Horizon is a fleet of six models from IFM, the model-building arm of MBZUAI, ranging from a 0.9B dense model to a 375B-A23B mixture-of-experts flagship.
  • The full technical paper landed on arXiv on 16 November 2026. The models first appeared in September; the paper is the detailed disclosure behind them.
  • Four models share one 22-trillion-token pre-training corpus and recipe. The flagship trained on 15.1 trillion tokens, of which roughly 10 trillion were synthetic.
  • New ideas in the paper include MoVA attention, softplus attention gating and Uno, a decoding speedup IFM puts at up to 2.2x.
  • Benchmark results are IFM’s own and have not been independently verified. The paper itself names expert-level scientific reasoning as the fleet’s main gap.

Most open model releases hand over the weights and stop there. K2-Horizon is built around the opposite argument: that a release is only genuinely open when the whole development history is public. That is the LLM360 standard IFM follows, and it is what makes this fleet worth paying attention to even in a month already crowded with frontier launches.

Six models, one shared recipe

The fleet covers the full range of deployment sizes. At the small end, the 0.9B dense model is aimed at memory-constrained devices such as phones. At the top, the 375B-A23B flagship is a multi-GPU or hosted-inference model for coding and agentic work. Between them sit 3.7B and 7B dense models, a 32B dense model and a 36B-A4B mixture-of-experts model. Four of the six share a single 22-trillion-token corpus and the same pre-training recipe, which makes the fleet a controlled experiment in scale and sparsity rather than six unrelated releases.

ModelTypeTotal parametersActive per tokenBest for
K2-Horizon 0.9BDense0.9B0.9BEdge devices, phones
K2-Horizon 3.7BDense3.7B3.7BSmall local deployments
K2-Horizon 7BDense7B7BGeneral local use
K2-Horizon 32BDense32B32BSerious coding and reasoning
K2-Horizon 36B-A4BMoE~37.4B~4.7BCost-efficient agentic work
K2-Horizon 375B-A23BMoE~379.2B~23.6BEnterprise coding and agents

Open science, not just open weights

The paper’s central claim is about transparency. For every model, IFM says it publishes the final weights, intermediate checkpoints from every training stage, pre-training data or the recipes that built it, fine-grained training logs, and the training, evaluation and infrastructure code — including its production-scale xLLM training framework, the RL360 reinforcement-learning stack and the Eval360 evaluation harness. Beyond the data itself, the team documents the pipelines that produced roughly 10 trillion synthetic pre-training tokens, the part of model development that open releases most rarely disclose.

That matters because it changes what other teams can do with the release. A lab can reproduce a training run, fine-tune from an intermediate checkpoint rather than the finished model, or audit exactly which data went into the model. The models and datasets are released under Apache 2.0, with the TxT360-v2 dataset under CC BY 4.0. One honest caveat: at launch, the two largest models shipped weights with some training data and code marked as forthcoming, so anyone building on the 375B or 36B models should check the repositories rather than take the openness claim on trust.

The new ideas: MoVA, softplus gating and Uno

On the modelling side, the paper introduces Mixture-of-Value Attention (MoVA), which applies mixture-of-experts to attention value projections, and softplus attention gating, a more flexible and theoretically motivated gate. The 36B-A4B model uses both and, according to IFM, nearly matches the 32B dense model trained on identical data — trailing it by only 1.2 to 3.7 points on MMLU, GSM8K, HellaSwag and HumanEval — while activating roughly one-seventh of the parameters. Uno, meanwhile, is a set of diffusion adapters that IFM says speeds up decoding by up to 2.2x without needing a separate draft model.

Benchmarks: IFM’s numbers, with a caveat

The headline results are IFM’s own, and no independent reproduction exists yet. On the numbers the team reports, the 375B-A23B flagship leads four recent open-weight rivals — GLM-5.2, MiniMax M3, Nemotron 3 Ultra and Inkling — on Toolathlon-Verified at 65.3 and on SWE-Atlas-QnA at 48.4, and it is reported to surpass GPT-5.6 Terra on Toolathlon-Verified. Its Terminal-Bench 2.1 headline of 70.2% came with an unusual piece of honesty: IFM re-audited every passing trial with a reward-hacking procedure and knocked 3.37 points off the score. The paper also names the gap itself: expert-level scientific reasoning, on benchmarks such as HLE and CritPt, is where the fleet still lags.

That kind of self-audit is rare, and it is exactly the sort of thing that becomes possible when the training and evaluation code are public. The results should still be read as the vendor’s until independent benchmarks arrive, but the paper gives outsiders the material to test them.

Where it fits in a crowded November

November has been a heavy month for model launches. Anthropic shipped Claude Fable 5.2 on 9 November, Google made Gemini 4 Carbon generally available on 10 November, and OpenAI released GPT-6.2 Astra on 17 November. Those are closed frontier models competing on capability and price. K2-Horizon is playing a different game: it is aimed at teams that want to inspect, reproduce and fine-tune a model rather than rent one through an API. For agentic workflows where a team needs to own the model — on-premise deployment, custom fine-tuning, or research into how a model was built — the fleet’s openness is the selling point, not a benchmark lead.

Availability and what to check first

Weights, checkpoints and datasets are published through IFM’s Hugging Face collections, with training and evaluation code under github.com/ifm-ai and an overview at ifm.ai/k2. The 375B flagship is a 758 GB download in full precision, so most users will reach it through a hosted inference partner or a quantised build rather than running it locally. Before building on the release, check the artifact repositories directly: confirm the training data and code for the two largest models have actually landed, and look at the intermediate checkpoints if you plan to fine-tune. The paper gives the recipe; the repositories are where you verify it.

What is K2-Horizon?

A fleet of six fully open language models from Abu Dhabi’s Institute of Foundation Models (IFM), ranging from a 0.9B dense model to a 375B-A23B mixture-of-experts flagship, built for reasoning, coding and agentic work.

When was the K2-Horizon technical paper published?

The full technical paper, arXiv:2611.21003, was submitted to arXiv on 16 November 2026. The models themselves first appeared in September 2026; the paper is the detailed disclosure behind the release.

What does ‘fully open’ mean for K2-Horizon?

Under the LLM360 definition, IFM releases not just the weights but intermediate checkpoints from every training stage, pre-training data or data-construction recipes, training logs, and all training, evaluation and infrastructure code.

How big is the K2-Horizon flagship model?

The 375B-A23B flagship has about 379 billion total parameters and activates roughly 23.6 billion per token. It was trained on 15.1 trillion tokens in two phases.

What are MoVA and Uno?

MoVA (Mixture-of-Value Attention) applies mixture-of-experts to attention value projections, and softplus attention gating is a more flexible gate design. Uno is a diffusion-adapter technique IFM says speeds up decoding by up to 2.2x without a separate draft model.

Where can I download the K2-Horizon models?

Weights and artifacts are published through IFM’s Hugging Face collections and the project page at ifm.ai/k2, with code under github.com/ifm-ai. All models are released under Apache 2.0.

Joseph Zada
Written by

Joseph Zada

Joseph is the founder and publisher of ChainAIWire, where he oversees editorial direction and strategic coverage at the intersection of blockchain infrastructure, digital assets, and artificial intelligence. Driven by a mission to bridge complex emerging technologies with clear, transparent reporting, Joseph tracks the convergence of decentralized protocols, autonomous AI systems, and evolving regulatory frameworks. Under his leadership, ChainAIWire delivers objective, high quality content reporting designed for developers, institutional allocators, and industry leaders navigating the modern digital economy.

More articles by this author →

Leave a Comment