Frontier Models
IFM at MBZUAI Releases K2 Horizon Fleet of Six Fully Open Models
The September 3, 2026 announcement provides six models from 0.9 billion to 375 billion parameters along with complete training data, code, checkpoints and evaluations under Apache 2.0 to support reproducibility.
K2 Horizon is a connected fleet of six fully open AI foundation models released by the Institute of Foundation Models at MBZUAI on September 3, 2026.
The Institute of Foundation Models at MBZUAI announced the K2 Horizon fleet on September 3, 2026.
The fleet consists of six models released under the Apache 2.0 license.
All components including weights, code, data and evaluations are provided on Hugging Face.
This approach supports full reproducibility of the results by external researchers.
What distinguishes the K2 Horizon release from earlier open model efforts?
Previous open releases often provided only model weights without additional artifacts.
K2 Horizon includes the full training data or detailed construction recipes.
The approach aligns with principles of open science as described by IFM leadership.
Reproducibility is enhanced by the inclusion of intermediate checkpoints and fine grained logs.
Post training artifacts further enable extension of the work by other teams.
Which models comprise the K2 Horizon suite and what are their specifications?
| Model | Parameters | Architecture | Notable Feature |
|---|---|---|---|
| 0.9B | 0.9 billion | Dense | Runs on watches and glasses with AIME 2026 above 48 |
| 3.7B | 3.7 billion | Dense | Strong performance at small scale on phones |
| 7B | 7 billion | Dense | State of the art results across reasoning and coding |
| 32B | 32 billion | Dense | High performance on mathematics benchmarks |
| 36B-A4B | 36 billion | MoVA | Uses Mixture of Value Attention for efficiency |
| 375B-A23B | 375 billion | MoE | Flagship model for enterprise reasoning tasks |
The models share core architecture, vocabulary and training methodology.
Native support exists for up to 524288 token context across the fleet.
Deployment tooling is available immediately on Hugging Face.
The 0.9B model sets new state of the art results at its scale.
The 3.7B and 7B models similarly lead on reasoning, mathematics, coding and agentic benchmarks.
How was the pre-training conducted for the K2 Horizon models?
Each model is pretrained on approximately 20 trillion tokens.
The training mixtures include nearly 17 percent explicit reasoning trajectories.
Approximately 10 trillion synthetic tokens form part of the data.
Innovations include Mixture of Value Attention for improved attention mechanisms.
Diffusion distillation delivers roughly 3 times inference speedup without quality loss.
Some variants used 22 trillion tokens in total pre training.
What components are included in the K2 Horizon release package?
- Final model weights for all six sizes
- Complete training code and scripts
- Training data or detailed construction recipes
- Intermediate checkpoints from training runs
- Fine grained training logs and evaluations
- Post training artifacts and agentic tools
The ordered list above outlines the full set of deliverables.
These elements enable independent reproduction of the published results.
Hugging Face hosts all resources with 22 items in the collection.
What benchmark performance do the K2 Horizon models demonstrate?
The smaller models outperform prior open models at equivalent parameter counts.
Reasoning and agentic capabilities receive particular emphasis in the evaluations.
The results are documented in the released evaluation files.
What deployment options exist for the K2 Horizon models?
The models are available immediately on Hugging Face.
Day zero support is provided by vLLM, SGLang and Ollama.
This support enables rapid integration into existing inference pipelines.
The context length of 524288 tokens supports long document and agent workflows.
What implications does the K2 Horizon release carry for the AI market and stakeholders?
Enterprise users gain access to a range of model sizes for different deployment scenarios.
Researchers can inspect the full training process to build upon the work.
The inclusion of agentic artifacts supports development of autonomous systems.
Closed model providers face increased competition at multiple scales.
The release may accelerate adoption of open models in production environments.
Academic labs benefit from the documented recipes for curriculum design.
How did IFM leaders describe the goals of the K2 Horizon release?
Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it. K2 Horizon delivers on that need. Every model in the fleet ships with its training data, recipe and evaluations. This is open science, and we believe it’s the best path forward for AI.Eric Xing, Founder of IFM, and President and University Professor of MBZUAI
Hector Liu emphasized the fleet approach over single model releases.
The director noted that every model competes with the best open models at its size.
The complete methodology accompanies each model in the fleet.
What developments can be expected following the K2 Horizon announcement?
Further fine tuning by the community is anticipated on the released artifacts.
New agentic applications may emerge from the provided post training tools.
Additional benchmarks and comparisons will likely appear in follow on research.
The open methodology may influence training practices at other organizations.
Updates to the models could appear as new data mixtures become available.
Integration with additional inference engines beyond the initial three is probable.
Frequently asked
What license applies to the K2 Horizon models and artifacts?
The K2 Horizon models and all associated artifacts are released under the Apache 2.0 license.
Where can developers access the K2 Horizon models and supporting resources?
The models, datasets and resources are hosted on Hugging Face with immediate support from vLLM, SGLang and Ollama.
What context length do the K2 Horizon models support?
The models provide native support for up to 524288 token context length.
Sources
- Institute of Foundation Models — The Institute of Foundation Models today introduced K2 Horizon, a new fleet of six AI foundation models ranging from 0.9 billion to 375 billion parameters. The new models are fully open—including model weights, code, training data and methodologies.
- Institute of Foundation Models — Today IFM is releasing K2 Horizon, a connected fleet of six models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B. Each model is pretrained on approximately 20 trillion tokens. K2 Horizon 0.9B achieves an AIME 2026 score above 48.
- Hugging Face — K2 Horizon models, datasets, and supporting resources • 22 items
- @grok — IFM / MBZUAI shipped K2 Horizon suite of six fully open models (0.9B to 375B-A23B, dense + MoE) under Apache 2.0, including weights, code, data, and evals on Hugging Face.