K2 Horizon

IFM's fully open family of six Apache 2.0 models, from 0.9B to 375B-A23B, with a 512K context and open code, data recipes and checkpoints.

In short. K2 Horizon is a set of six open models from the Institute of Foundation Models at MBZUAI in Abu Dhabi, from a 0.9B model for edge devices to a 375B mixture-of-experts flagship. All are Apache 2.0, read up to 512K tokens (128K for the smallest), and come with training code, data or data recipes, and intermediate checkpoints. They are built for reasoning, coding and agent tasks.

Summary

IFM, the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and the team behind LLM360, released K2 Horizon on 3 September 2026: six models named 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B. The 0.9B, 3.7B, 7B and 32B are dense, the 375B-A23B is a sparse mixture of experts with 23B active parameters, and the 36B-A4B (about 4B active) uses IFM's new Mixture-of-Value Attention, MoVA, which routes each token through 4 of 64 value projections inside attention without enlarging the KV cache. IFM says each model saw roughly 20 trillion training tokens, that nearly 17% of the pretraining corpus is problem-solving trajectories with explicit reasoning, and that about 10 trillion tokens are synthetic. The context window is 524,288 tokens at every size except the 0.9B, which has 131,072. The Hugging Face parameter counts include embeddings and are larger than the names suggest: 9.0B for the 7B, 34.8B for the 32B, 37.4B for the 36B-A4B and 379.2B for the 375B-A23B.

IFM calls the release fully open: weights, training code, data or data recipes, intermediate checkpoints from every stage and training logs, with the models and code under Apache 2.0 and the datasets under their own licenses, such as ODC-BY. The model cards, last updated on 1 October 2026, still listed the technical report as in progress and due by the end of September 2026, and we did not find one. In IFM's tables the 375B-A23B scores 70.2 on Terminal-Bench 2.1 (Nemotron 3 Ultra 53.9, MiniMax-M3 65.2, GLM 5.2 77.9), 72.8 on BrowseComp (Inkling 77.1, MiniMax-M3 83.5) and 87.3 on GPQA Diamond (MiniMax-M3 92.9). The 36B-A4B scores 58.6 on Terminal-Bench 2.1, against 44.9 for Qwen3.6-35B-A3B and 51.7 for Muse Glimmer-30B, but trails them on GPQA Diamond. The 32B trails Qwen3.8-27B in its own table, 62.6 against 79.8 on Terminal-Bench 2.1 and 24.1 against 48.0 on tau3-Banking, and the 7B reports 72.4 on SWE-bench Verified against 50.8 for Qwen3.5-9B. All scores are IFM's own.

The BF16 files are 758.3 GB for the 375B, 74.9 GB for the 36B-A4B, 69.6 GB for the 32B, 18.0 GB for the 7B, 10.1 GB for the 3.7B and 2.2 GB for the 0.9B. IFM also publishes FP8 (390.5 GB) and NVFP4 builds of the 375B, and GGUF files for every other size, with Q4_K_M at 22.4 GB for the 36B-A4B, 21.1 GB for the 32B, 5.6 GB for the 7B, 3.2 GB for the 3.7B and 0.67 GB for the 0.9B. The GGUF cards say you need a llama.cpp release with K2 Horizon support, b11454 or later. The vLLM and SGLang commands use trust_remote_code and K2 Horizon reasoning and tool-call parsers, and IFM recommends high reasoning effort, temperature 1.0, top-p 0.95 and at least 32,768 output tokens. Uno, a LoRA adapter trained with diffusion distillation that IFM says speeds up generation about threefold, is published for the 7B and 0.9B. In the 30 days to 11 November 2026, Hugging Face counted about 54,700 downloads of the 7B, 17,100 of the 3.7B, 14,900 of the 36B-A4B and 10,100 of the 375B.

Good for

  • Researchers and teams that want a strong open model with its training code, data recipes, checkpoints and logs, to reproduce or adapt reasoning and agent training
  • Long-context agents, since every size except the 0.9B takes 512K tokens
  • A local coding agent on a workstation, using the 36B-A4B at 22.4 GB in 4-bit or the 7B at 5.6 GB
  • Prototyping small and scaling up within one family that shares its architecture, vocabulary and tooling

Watch out for

  • Every score is IFM's own, and in IFM's table Qwen3.8-27B is well ahead of the 32B on Terminal-Bench 2.1, tau3-Banking and GPQA Diamond
  • No technical report was available when we checked, although IFM's cards promised one by the end of September 2026
  • The 375B needs an 8-GPU node in BF16, loads with trust_remote_code and IFM's parsers, and GGUF files need a recent llama.cpp
  • The datasets keep their own licenses, and some are described by recipe rather than redistributed
  • It handles text only

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the 0.9B (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
0.9B (BF16) BF16 1.08B 2.48 GB 1.24 GB 0.745 GB Laptop or phone
3.7B (BF16) BF16 5.06B 12 GB 5.82 GB 3.49 GB Laptop or phone
7B (BF16) BF16 9B 21 GB 10 GB 6.21 GB Laptop or phone
7B GGUF Q4_K_M INT4 9B 21 GB 10 GB 6.16 GB Laptop or phone
32B (BF16) BF16 34.8B 80 GB 40 GB 24 GB Gaming GPU
MoVA-36B-A4B (BF16) BF16 37.4B 4B active 86 GB 43 GB 26 GB Gaming GPU
MoVA-36B-A4B GGUF Q4_K_M INT4 37.4B 4B active 86 GB 43 GB 25 GB Gaming GPU
375B-A23B (BF16) BF16 379B 23B active 872 GB 436 GB 262 GB Big-memory desktop
375B-A23B (FP8) FP8 379B 23B active 872 GB 436 GB 262 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is K2 Horizon?
K2 Horizon is a set of six open models from the Institute of Foundation Models at MBZUAI in Abu Dhabi, from a 0.9B model for edge devices to a 375B mixture-of-experts flagship. All are Apache 2.0, read up to 512K tokens (128K for the smallest), and come with training code, data or data recipes, and intermediate checkpoints. They are built for reasoning, coding and agent tasks.
Can I use K2 Horizon commercially?
Yes. K2 Horizon is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is K2 Horizon open source?
K2 Horizon is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run K2 Horizon on my own computer?
Yes. The smallest variant, 0.9B (BF16), needs about 0.745 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was K2 Horizon released?
K2 Horizon was released by Institute of Foundation Models (MBZUAI) on September 3, 2026.
How long a context window does K2 Horizon have?
K2 Horizon supports a context window of up to 512K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Jais 3 Institute of Foundation Models (MBZUAI) 1.08B – 379B 23B active 512K Apache 2.0 (Permissive) 0.745 GB Laptop or phone
Nemotron 3 Ultra NVIDIA 550B 55B active 1M OpenMDW-1.1 (Permissive) 379 GB Big-memory desktop
MiniMax M3 MiniMax 428B 23B active 1M MiniMax Community (Conditional) 295 GB Big-memory desktop
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU