K2-Horizon-7B-GGUF

K2-Horizon-7B-GGUF

Quantized GGUF copies of IFM's K2-Horizon-7B text model for running locally in llama.cpp, from BF16 down to 4-bit, with 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: This repository holds GGUF copies of K2-Horizon-7B, a decoder-only text model from IFM with a 7B core and a 512K-token context window. The GGUF format lets you run it locally with llama.cpp instead of calling a hosted service.

Each file carries its tokenizer data and a chat template that llama.cpp can read, so a chat session needs no extra setup files.

Who it suits: People who want to run an open 7B-class model on their own computer and read long inputs. The card says IFM tested the base model on agentic, coding, long-context and reasoning benchmarks.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local coding and tool-use model | A coding helper that runs offline on a workstation
Research | ML researchers | Study an open model and its released checkpoints | Comparing behavior across training stages
Documents | Analysts | Feed long texts into one prompt | Asking questions about a long report
Hobbyists | Local AI users | Try a quantized model on modest hardware | A 4-bit copy on a home PC

You can skip it if you want a hosted chat service, or if your llama.cpp build is older than the release the card names.

Where to use it: The files work with llama.cpp and tools built on it, on Windows, macOS and Linux.

• llama.cpp, built from a release that supports the K2 Horizon architecture
• Windows, macOS and Linux, where you can run llama.cpp
• Any chat front end that loads GGUF files through llama.cpp

Use case | Who | What they get back
Local chat | Anyone with a recent PC | Replies from the model without a cloud account
Long-document questions | Analysts and writers | Answers drawn from a lengthy input
Coding and agent tests | Developers | A local model to try on agentic and coding tasks

How to set it up: You need Python for the download tool and a llama.cpp build that knows the K2 Horizon architecture. The card says mainline support arrived in release b11454.

• Python with pip
• A llama.cpp build from release b11454 or newer
• Disk space: about 4.5 GB for the 4-bit copy, an estimate from Hugging Face

1. Install the Hugging Face hub tools with pip.
2. Download the repository files with the hf command.
3. Point llama.cpp at the GGUF variant you want.

pip install -U huggingface_hub
hf download IFM/K2-Horizon-7B-GGUF

How to use it: Load one of the GGUF files in llama.cpp and chat with it. The built-in chat template means the prompt format is handled for you. The card lists these variants.

Variant | What it is
BF16 | The full-precision copy, the largest file
Q8_0 | An 8-bit quantization
Q6_K | A 6-bit quantization
Q5_K_M | A 5-bit quantization, mixed
Q5_0 | A 5-bit quantization, plain
Q4_K_M | A 4-bit quantization, the smallest listed

Example: You open the Q4_K_M file in llama.cpp and ask it to explain a stack trace from your project. The model replies in the chat, running entirely on your own machine.

For newcomers and experts: If you are new, pick the 4-bit Q4_K_M file first, since it is the smallest and the easiest on memory. Use a recent llama.cpp, load the file and type a question. If the answers feel weak, try a larger variant.

If you know the field: K2-Horizon-7B sits in the middle of the K2-Horizon family and is a dense model. Native 524,288-token context starts from the midtraining stages. IFM released intermediate checkpoints, plus training data, recipe, code and evaluation resources. Lower quantization shrinks the file and costs some accuracy.

Download a GGUF file from the repository
Load it in a llama.cpp build with K2 Horizon support
The built-in chat template formats your prompt
The model replies on your own machine

Good to know: The card's table compares the model with Gemma 4-12B, Qwen3.5-9B and Granite 4.2-8B on competition mathematics. Treat these as the authors' own numbers.

Title: HMMT Feb 2026 competition math score
K2-Horizon-7B: 73.3 points
Granite 4.2-8B: 66.5 points
Qwen3.5-9B: 65.7 points
Gemma 4-12B: 63.1 points

• License: Apache 2.0
• Cost to run: none named, you run it on your own hardware
• Data it reaches: the files you download from Hugging Face and the prompts you type locally
• Not for: older llama.cpp builds without K2 Horizon support

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 9B
• Weights: about 4.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 24,083
• Likes: 84
• License: apache-2.0
• Languages: en
• Based on: IFM/K2-Horizon-7B
• Fine-tunes: 1
• Tags: transformers, gguf, k2-horizon, 7b, dense, open-weights, text-generation, conversational

Links

• Model: https://huggingface.co/IFM/K2-Horizon-7B-GGUF
• Files: https://huggingface.co/IFM/K2-Horizon-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-7B-GGUF

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