K2 Horizon

K2 Horizon is a family of fully open agentic models from the Institute of Foundation Models (IFM), released under the Apache-2.0 license. The models reason before they answer, call tools, and handle up to 512K tokens of context, at sizes from 0.9B to 36B parameters.

Highlights

• Agentic. Built for tool use, coding, and multi-step tasks, with native tool calling.
• Reasoning. Every model thinks before answering, and the reasoning is returned separately from the answer.
• 512K context. Native 524,288-token context window (128K for 0.9B).
• Efficient mixture of experts. K2-Horizon-MoVA-36B-A4B activates only 4B parameters per token.
• Fully open. Training data and recipe, training code, evaluation resources, and intermediate checkpoints are public or being released.
• Apache 2.0. Free for commercial and non-commercial use.

Models

Every size is also available as q4_K_M, q5_0, q5_K_M, q6_K, q8_0 and bf16, for example ifm/k2-horizon:7b-q8_0. The smaller sizes default to q8_0 because 4-bit quantization costs them noticeably more quality; for 7b and up, q4_K_M scores within a few points of bf16 on our evaluations.

Quantization quality

How each quantization compares with full-precision bf16 is on each GGUF model card: 0.9B 3.7B 7B 32B MoVA-36B-A4B

Get started

Requires Ollama 0.40.4 or later.

ollama run ifm/k2-horizon

The reasoning is returned in message.thinking, the answer in message.content, and tool calls in message.tool_calls.

Note: Ollama doesn’t yet pass the model’s reasoning from earlier turns back to it in multi-turn chats and tool loops. A fix is under review in ollama/ollama#18281.

Best practices

1. Reasoning effort: high. K2 Horizon reasons at high effort by default, which is what all reported results use. Reasoning can’t be turned off.
2. Sampling. Temperature and top_p are preset from each model’s card: temperature 1.0 and top_p 0.95, or 0.6 and 0.95 for 0.9b.
3. Context length. Ollama picks a default context size from your GPU memory, which is 4,096 tokens on GPUs with less than 24 GB. Reasoning, coding and agent tasks need much more, so set at least 32,768:

   OLLAMA_CONTEXT_LENGTH=32768 ollama serve

or pass "options": {"num_ctx": 32768} in API requests. Allow enough room for the reasoning: a response cut off mid-reasoning has no answer.

Model details

• Developer: Institute of Foundation Models (IFM)
• Language: English
• License: Apache 2.0

Reference

• K2 Horizon GGUF collection
• K2 Horizon Collection
• Announcement: Introducing K2 Horizon

@misc{k2horizon2026,
  title  = {Introducing K2 Horizon: Frontier Performance, Radically Open},
  author = {{IFM Team}},
  year   = {2026},
  url    = {https://ifm.ai/blog/k2/},
}