K2 Horizon: Redefining the Frontier of Open Performance

Date: September 3, 2026
Organization: Institute of Foundation Models (IFM)
Tags: #AI #OpenSource #LLM #K2Horizon

Today, the IFM is proud to unveil K2 Horizon, a sophisticated ecosystem consisting of six interconnected models. This fleet is designed to push the boundaries of what is possible across various scales, from ultra-compact edge devices to massive enterprise infrastructures.

The fleet includes the following models:

• $375\text{B-A23B}$
• $36\text{B-A4B}$
• $32\text{B}$
• $7\text{B}$
• $3.7\text{B}$
• $0.9\text{B}$

Across the board—specifically in mathematics, coding, reasoning, agentic workflows, and general utility—K2 Horizon delivers elite performance. Notably, the $0.9\text{B}$, $3.7\text{B}$, and $7\text{B}$ variants establish a new state-of-the-art (SOTA) for their respective size classes.

---

🔓 A New Standard for Radical Openness

K2 Horizon isn't just a release of weights; it is the most exhaustive open-source effort to date. We are peeling back the curtain on the entire training lifecycle, moving from initial pretraining through to reasoning and agentic post-training.

What is being released?

To ensure total transparency and reproducibility, we are providing:

• Final Model Weights
• Intermediate Checkpoints (across all stages)
• Training Data (or comprehensive construction recipes)
• Open Architecture & Mixture Compositions
• Training Code & Configurations
• Fine-grained Logs & Evaluation Results

Licensing Note: Models and code are provided under the Apache 2.0 license. Datasets are released under applicable licenses (e.g., ODC-BY). In cases where redistribution is prohibited, we provide the exact recipes used to construct and mix the data.

The Development Pipeline

The following diagram illustrates the transparency of the K2 Horizon lifecycle:

---

🚀 Performance Across the Spectrum

K2 Horizon provides a versatile toolkit for any deployment scenario. By maintaining a consistent core architecture, vocabulary, and interface (with the exception of a streamlined vocabulary for the $0.9\text{B}$ model), users can easily route tasks between different model sizes.

Model Capabilities Matrix

Technical Highlights

• The MoVA Advantage: The $36\text{B-A4B}$ model utilizes a novel Mixture-of-Value-Attention (MoVA) mechanism. This allows it to punch well above its weight, outperforming models significantly larger than itself in terms of active parameter efficiency.
• Small-Scale Power: The $0.9\text{B}$ model is a powerhouse for its size, achieving an AIME 2026 score exceeding 48, while maintaining strong tool-use capabilities.
• Agentic Proficiency: The $3.7\text{B}$ and $7\text{B}$ models excel in multi-step environments, as evidenced by their performance on BrowseComp and SWE-bench.
• The Ceiling: While TerminalBench tasks (which require exhaustive exploration and recovery) remain a challenge for the smallest models, K2 Horizon pushes the absolute limit of what small-scale AI can achieve.

---

💡 Why the "Horizon" Approach Matters

We believe that a transparent model is useless if it is obsolete, and a powerful model is a "black box" if only the weights are shared.

The K2 Horizon philosophy bridges this gap:

1. Competitive Power: We provide models that actually compete at the frontier.
2. Educational Value: By releasing the "recipes," researchers can study how reasoning and planning emerge rather than treating the model as an opaque object.

This commitment stems from our 2023 LLM360 paper. We have evolved from releasing simple open models to providing the complete lifecycle of agentic post-training.

Example Configuration Snippet

For those looking to reproduce our results, the released configurations follow a structured format:

model_architecture: "K2-Horizon-Sparse"
mixture_of_experts:
  total_params: 375B
  active_params: 23B
  attention_mechanism: "MoVA" # For 36B-A4B variant
training_stage: "agentic_post_training"
optimizer: "AdamW"
learning_rate: 1.2e-5

---

🔍 Deep Dive: K2 Horizon 375B-A23B

The 375B-A23B is the crown jewel of the fleet, designed for the most demanding enterprise applications.

• Architecture: It employs a sparse Mixture-of-Experts (MoE) design.
• Efficiency: While it possesses a total capacity of $375\text{B}$ parameters, it only activates $\approx 23\text{B}$ parameters per token.
• Positioning: It ranks as one of the top-performing models in the sub-400B parameter category, specifically dominating in complex reasoning, software engineering, and agentic evaluations.