K2-Horizon-7B-Uno

K2-Horizon-7B-Uno

A LoRA adapter that adds a diffusion pathway to IFM's K2-Horizon-7B language model, for research into faster text generation. It holds only the adapter, not the base weights.

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-7B-Uno is a diffusion-augmented language model from IFM, described in a paper on lossless speedups for LLMs through discrete diffusion. It has two pathways in one architecture.

The autoregressive pathway reuses the weights of K2-Horizon-7B. The diffusion pathway adds LoRA-based diffusion adapters on top of those weights. This repository holds only the adapter, and the base weights live in a separate repository.

Who it suits: Researchers and engineers who study faster text generation and can work with a PEFT adapter on top of a base model. It is not a ready-to-chat product.

Industry | Who uses it | How they use it | Example
Research | ML researchers | Study diffusion-based decoding next to autoregressive decoding | Comparing the two pathways on the card's benchmarks
Software | Inference engineers | Test whether adapter-based speedups fit a serving stack | Measuring speed and score on their own tasks
Agents | Agent developers | Try the model on tool-use benchmarks | Running the card's tau-style tool-use evaluations

You can skip it if you want a model that works from a single download or a plain chat model, since this needs the base model as well.

Where to use it: It runs on Windows, macOS or Linux wherever Python and the PEFT library work. The card links to a project page and a code repository for the research.

• Python environments with PEFT and the base model
• Research notebooks and evaluation scripts
• The project's own evaluation scripts, a selection of which is public

Use case | Who | What they get back
Reproduce results | Researchers | The adapter to load over the base model
Speed experiments | Inference engineers | A diffusion pathway to compare with normal decoding
Benchmark runs | Agent developers | Scores on tool-use and other tests listed in the card

How to set it up: You need Python and the Hugging Face tools. Because the repository contains only the adapter, you also need the base model, K2-Horizon-7B, from its own repository.

• Python with pip
• The base model, IFM/K2-Horizon-7B, downloaded separately
• PEFT to attach the adapter, since the library is listed as peft

1. Install the Hugging Face hub tools with pip.
2. Download the adapter files with the hf command.
3. Download the base model the same way and load the adapter over it.

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

How to use it: The visible part of the card gives no run command. It points to the project's code, where selected evaluation scripts are public and the full suite is planned for later. Use those scripts to run the model.

Example: A researcher downloads the adapter and the base model, then runs one of the published evaluation scripts on a benchmark. The output is a benchmark score, which the card pairs with a speed figure called TPF.

For newcomers and experts: If you are new, know that this is a small add-on file, not a whole model. It does nothing alone. Start by downloading the base model and the adapter, then read the project page to see how they are combined.

If you know the field, the design keeps the base weights for the autoregressive pathway and trains LoRA diffusion adapters for a second pathway. The card's table lists a score and a TPF figure per benchmark and compares Uno at 8B with Nemotron-Diff, Mercury 2 and Diff-Gemma.

Load base K2-Horizon-7B weights
Attach the Uno LoRA adapter
Choose the autoregressive or diffusion path
Generate text and score it on a benchmark

Good to know: The card's evaluation table shows the Uno adapter ahead of the listed diffusion models on the tool-use telecom test, as the chart shows. Some cells in the card are marked unavailable.

Title: Telecom tool-use benchmark score
Uno: 90.1 points
Mercury 2: 71 points
Diff-Gemma: 68.1 points
Nemotron-Diff: 14.3 points

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the adapter and base model files you download from Hugging Face
• Not for: standalone use, since it holds only the adapter and needs the base model

On Hugging Face, 13 November 2026

• Task: text generation
• Library: peft
• Formats: Safetensors
• Downloads in the last 30 days: 203
• Likes: 113
• License: apache-2.0
• Based on: IFM/K2-Horizon-7B
• Quantized versions: 3
• Fine-tunes: 1
• Tags: peft, safetensors, lora, diffusion-language-model, conditional-lora, text-generation

Links

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

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