K2-Horizon-3.7B

K2-Horizon-3.7B

Handles coding, reasoning and agent-style tasks as a small open model with a 512K token context window. Built by IFM for people who want public training data, code and mid-training checkpoints.

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 is IFM's small dense language model, one of several in its K2-Horizon line. It reads text and writes text with a core of 3.7B parameters and a 512K context window. The team has made its training data, recipe, code and evaluation resources public.

Who it suits: It suits people who want a small open model for coding and reasoning work, and researchers who want to study how a model's abilities develop during training.

Industry | Who uses it | How they use it | Example
Research | Language model researchers | Study how abilities change across training using the released checkpoints | Compare an early checkpoint's answers with the final model's
Software | Developers | Try a small model on coding and reasoning tasks | Send it a short function and ask what it does
Education | Students and teachers | Learn from a model whose data and code are open | Follow the public training recipe and evaluation resources

People who need a much larger model can skip it, and so can anyone who needs languages other than English.

Where to use it: The card excerpt names no app. The files load in any Python project that can read them.

• Windows, macOS and Linux, with Python for the download
• Python projects built on the transformers library, which the record lists

Use case | Who | What they get back
Coding and reasoning questions | Developers and students | A written answer from a small dense model
Long inputs | Anyone with large documents | A reply that can draw on a window of 524,288 tokens

How to set it up: You need Python with pip, and room for the weights.

• Python with pip
• About 10.1 GB of disk for the BF16 weights, or about 2.5 GB at 4-bit, which is an estimate

1. Install the Hugging Face hub library with pip.
2. Download the repository's files to your computer with the hf command.
3. Load the downloaded files with the transformers library in your own script.

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

How to use it: The card excerpt shows no run command, so use the transformers library the record lists. Load the downloaded model and its tokenizer, send a prompt and read the reply. The 524,288-token window is built in from the middle training stages on.

Example: Paste in a short function and ask the model to explain what it does and where it could fail. The card says it was tested on agentic, coding and reasoning benchmarks, so a coding question fits what it was built for.

For newcomers and experts: If you are new to this, download the files with the two commands above and load them with transformers. Try one short coding question first. Hugging Face lists 5.1B parameters, while the card describes a 3.7B core.

If you know the field, this is a dense decoder-only model with a long native context. The team also releases checkpoints from partway through training, so you can watch abilities change over time instead of seeing only the final model.

You send a prompt
The dense decoder-only model reads it, up to 512K tokens
The model writes its answer

Good to know: The card's comparison table lists the parameter counts below next to other open dense models.

• License: Apache 2.0, as the Hugging Face record states
• Languages: English only in the record
• Other versions: the record counts 26 quantized copies and 11 fine-tuned ones made by other people
• Cost to run: none named
• Data it reaches: the model files you download from Hugging Face, plus the prompts you send it

Title: Parameters in the card's comparison table, in billions
K2-Horizon-3.7B: 3.7 billion
Qwen3.5-4B: 4 billion
G9v3-3B: 3 billion
Granite 4.2-3B: 3 billion
Nemotron 3 Nano-4B: 4 billion

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 5.1B
• Weights: about 10.1 GB in BF16; about 2.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 15,489
• Likes: 115
• License: apache-2.0
• Languages: en
• Quantized versions: 26
• Fine-tunes: 10
• Tags: transformers, safetensors, k2_horizon, text-generation, k2-horizon, 3.7b, dense, open-weights, conversational

Links

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

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