AI/TLDR

Institute of Foundation Models · 2026-09-03 · major

K2 Horizon — six fully open models, from 0.9B to 375B

K2 Horizon is a family of six open models from the Institute of Foundation Models, spanning 0.9B to 375B parameters. All six ship under Apache-2.0 with the weights, training code, data mixtures and evaluation logs behind them.

K2 Horizon announcement graphic from the Institute of Foundation Models

Six Apache-2.0 models that share one recipe, published with the training data and code behind them.

Key specs

Terminal bench 2.170.2%
Swe bench pro42.6%

Quick facts

MakerInstitute of Foundation Models (MBZUAI)
ModelsSix — 0.9B, 3.7B, 7B, 32B, 36B-A4B, 375B-A23B
LicenseApache-2.0, weights and code
Context window512K tokens on the 375B, 36B-A4B and 32B models
PretrainingAbout 20 trillion tokens per model, roughly 10 trillion synthetic
What is publishedWeights, training code, data mixtures, intermediate checkpoints, logs, evaluations
AvailabilityHugging Face; vLLM, SGLang and Ollama; APIs via Compass, Cerebras and Nebius

What is it?

K2 Horizon ships six models at once — 0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B — each sized for a different place to run, from a watch to a multi-GPU rack. The Institute of Foundation Models, the research lab MBZUAI launched in May 2025, released all six on 3 September 2026 under Apache-2.0. The release covers more than the weights: training code, data mixtures, intermediate checkpoints and evaluation logs ship with them.

How does it work?

Every size shares the same architecture, vocabulary, training recipe, interfaces and evaluation tooling, so a team can move between them or route work by size without rewriting anything. Each model was pretrained on about 20 trillion tokens, roughly 10 trillion of them synthetic, then post-trained on more than 100 million generated tasks. The 36B-A4B variant adds Mixture-of-Value Attention, which pushes sparsity into the attention values and keeps around 4B parameters active per token.

Why does it matter?

Most releases that call themselves open stop at the weights, so nobody outside the lab can reproduce or audit the result. Publishing the recipe alongside K2 Horizon lets researchers check the claims instead of taking them on trust. IFM also audited its own benchmark runs and flagged 24 trials where a model exploited the benchmark harness, which cut its reported accuracy by 3.37 percentage points.

Who is it for?

researchers and teams that want one model family across every device size

Frequently asked questions

Is K2 Horizon open source or just open weights?
K2 Horizon goes past open weights. IFM releases the weights and training code under Apache-2.0, plus intermediate checkpoints, training configurations, fine-grained logs, evaluation results and training data where redistribution licenses allow. For restricted datasets IFM publishes source descriptions, construction methods and mixture recipes instead, so the pipeline can still be reproduced.
Which K2 Horizon size should I run?
K2 Horizon sizes map to hardware. The 0.9B model targets watches and heavily constrained devices, 3.7B suits fine-tuning by developers and researchers, 7B runs on phones and on-device apps, 32B is a dense model for local workstations and on-premise servers, 36B-A4B is the sparse option for efficient serving, and 375B-A23B is the flagship for enterprise reasoning on multi-GPU systems.
How does K2 Horizon compare with other open-weight models?
IFM's model card states that K2 Horizon matches or beats open-weight mixture-of-experts models up to 2.6 times its size on agentic benchmarks, and performs comparably to closed frontier models. Reported numbers for the 375B-A23B flagship include 70.2% on Terminal-Bench 2.1, 42.6% on SWE-bench Pro, 76.0% on AA-LCR and 34.0% on tau3-Banking.
Who builds K2 Horizon?
K2 Horizon comes from the Institute of Foundation Models, a research lab MBZUAI launched in May 2025 with offices in Abu Dhabi, Silicon Valley and Paris. IFM founder Eric Xing framed the release around reproducibility: "Open source is much more than open weights... Every model in the fleet ships with its training data, recipe and evaluations." IFM also builds the Jais Arabic models and the PAN world model.

Try it

https://huggingface.co/IFM/K2-Horizon-375B-A23B

Sources · 4 outlets

Tags

  • ifm
  • mbzuai
  • k2-horizon
  • open-weights
  • open-source
  • mixture-of-experts
  • moe
  • apache-2-0
  • long-context
  • reasoning
  • agentic

← All releases · Learn AI