K2-Horizon-3.7B-GGUF

K2-Horizon-3.7B-GGUF

GGUF files of the 3.7B K2-Horizon text model in six precision levels, from BF16 to Q4_K_M, for running with llama.cpp. The model has a 512K-token context window.

Made by IFM on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: K2-Horizon-3.7B-GGUF holds GGUF versions of IFM's K2-Horizon-3.7B, a small dense decoder-only text model. The files are meant for llama.cpp and come in six precision levels.

Each file carries tokenizer data and a chat template that llama.cpp can read. The base model has a 512K context window and was released with its training data, recipe, code and evaluation resources.

Who it suits: Developers and researchers who want a small open model they can run on their own hardware. It also suits students who want to study a model whose training material is public.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a small text model locally through llama.cpp | Load the Q4_K_M file for a light assistant
Research | Training researchers | Study how abilities change across training stages | Look at the intermediate checkpoints of the base model
Education | Students | Learn how an open model is built from public material | Read the published recipe and evaluation resources
Agent building | Prototype builders | Try a small model on coding and agent style prompts | Test it on your own coding questions

Anyone who wants a finished chat app can skip it, since these are model files that you load in llama.cpp yourself.

Where to use it:

• llama.cpp, on a build that supports the K2 Horizon architecture
• Windows, macOS or Linux with Python, where the download steps run
• Your own computer, since the 4-bit weights are estimated at about 2.5 GB

Use case | Who | What they get back
Run a small local assistant | Developers | Text replies from a 3.7B model on your own machine
Try coding and agent prompts | Developers | Output from a model the card tested on agentic, coding and reasoning benchmarks
Work with long input | Analysts | Room for a context window of 524,288 tokens

How to set it up: Check your llama.cpp version first, because older builds cannot load this architecture.

• Python and pip, for the Hugging Face download tools
• A llama.cpp build from release b11454 or later, which is where mainline support for K2 Horizon was merged
• Disk space for the variant you pick

1. Install the huggingface_hub package.
2. Download the repository with the hf command, or pick a single variant file from its Files tab.
3. Update llama.cpp to a build with K2 Horizon support.
4. Load the GGUF file in llama.cpp.

pip install -U huggingface_hub
hf download IFM/K2-Horizon-3.7B-GGUF

How to use it: Load one of the GGUF files in llama.cpp and chat with it. Each file already holds the tokenizer data and the chat template, so you do not supply them. Pick the variant by how much memory you have.

File variant | What the name tells you
BF16 | The original BF16 precision, with no quantization
Q8_0 | 8-bit quantization
Q6_K | 6-bit quantization
Q5_K_M | 5-bit quantization
Q5_0 | 5-bit quantization, another scheme
Q4_K_M | 4-bit quantization, the smallest variant listed

Example: You load the Q4_K_M file in a llama.cpp build with K2 Horizon support. You type "Write a Python function that parses a CSV file." The chat template in the file formats it, and the model answers in text.

For newcomers and experts: Think of these files as the same small model saved at different sizes. Smaller files use less memory and lose some precision. A good first try is the Q4_K_M file, since it is the smallest listed.

Underneath, the base model is a dense decoder-only network with a 3.7B core. It handles 524,288 tokens of context, a length it was trained for from the midtraining stages. The trade-off with quantization is lower precision for less memory. The card compares it with other open dense models of 3B to 4B parameters.

Download a GGUF variant file
Update llama.cpp to a supporting build
Load the file with its built-in template
Chat with the model

Good to know: The card shows a benchmark chart as an image, so its scores are not repeated here.

Title: Parameters in billions
K2-Horizon-3.7B: 3.7 B
Qwen3.5-4B: 4 B
G9v3-3B: 3 B
Granite 4.2-3B: 3 B
Nemotron 3 Nano-4B: 4 B

• Size: the card describes a 3.7B core, and Hugging Face's record lists 5.1B parameters for these files
• Memory: the 4-bit files are estimated near 2.5 GB
• Language: English is the one listed
• License: Apache 2.0
• Cost to run: no API key or paid service is named
• Data it reaches: the files you download and the prompts you type into llama.cpp on your own machine
• Not for: the card names no restrictions beyond its license

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 5.1B
• Weights: about 2.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 19,581
• Likes: 60
• License: apache-2.0
• Languages: en
• Based on: IFM/K2-Horizon-3.7B
• Tags: transformers, gguf, k2-horizon, 3.7b, dense, open-weights, text-generation, conversational

Links

• Model: https://huggingface.co/IFM/K2-Horizon-3.7B-GGUF
• Files: https://huggingface.co/IFM/K2-Horizon-3.7B-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download IFM/K2-Horizon-3.7B-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.