K2 Horizon: Six Fully Open AI Models, With Training Data and Code Included
Abu Dhabi's Institute of Foundation Models has released K2 Horizon, a family of six language models ranging from 0.9 to 375 billion parameters. You don't just get the models: the training data, intermediate checkpoints and the full codebase are all up for download too. The technical paper is now out as well.

What's going on
Abu Dhabi's Institute of Foundation Models (IFM) has released an entire family of language models, which it calls a "fleet": K2 Horizon. It includes six models, from a tiny 0.9B model that the IFM says can even run on a smartwatch up to a 375-billion-parameter flagship. According to its press release, the IFM first unveiled the models on September 3, 2026. Since November 16, the detailed technical paper, "K2-Horizon: A Frontier Fleet of Open Source Language Models," has also been available on Hugging Face Papers.
The real story isn't performance. It's how open the project is. Many labs only publish the finished weights ("open weights"). The IFM goes much further: it also releases the training data, checkpoints from every training phase, training logs, evaluation code and the complete training infrastructure. You'll find everything on the project page at ifm.ai/k2. The models are hosted on Hugging Face, and weights and code are licensed under Apache 2.0.
The IFM is backed by the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), which founded the institute in May 2025. It now has offices in Silicon Valley, Paris and Abu Dhabi. Its founder is Eric Xing, who is also president of MBZUAI. The IFM's portfolio also includes the Arabic language model Jais and the world model PAN.
The six models at a glance
All models share the same architecture, vocabulary, training method and tooling. Only the smallest model has a smaller vocabulary. Two of them are Mixture-of-Experts (MoE) models. In this design, only a small part of the network works on each word, much like a large company where only the relevant specialists handle a given job. The label "A23B" means that of 375 billion total parameters, only 23 billion are active per token.
| Model | Architecture | Intended use (per IFM) |
|---|---|---|
| K2 Horizon 0.9B | dense | watches, glasses, very low-end hardware |
| K2 Horizon 3.7B | dense | smartphones, good for fine-tuning |
| K2 Horizon 7B | dense | smartphones, software development, research |
| K2 Horizon 32B | dense | laptops and your own servers |
| K2 Horizon 36B-A4B | MoE, 4B active | cheap to run on your own hardware |
| K2 Horizon 375B-A23B | MoE, 23B active | flagship for agents and enterprises |
According to the IFM, the 0.9B, 3.7B and 7B models are each the best in their size class at math, reasoning and tool use. It calls the 7B model the best model under 10 billion parameters. The paper is more cautious: it says the models lead or match strong open models of the same size on agent and coding benchmarks. The available sources don't include concrete benchmark numbers, so these claims can't be verified here.
The IFM's dynamic model routing is also built with practical use in mind. Each task is automatically sent to the cheapest model that can handle it. The idea is that developers can start experimenting with small models and later move to the big ones without rebuilding their workflows.
What's new under the hood: MoVA and Uno
The paper introduces several new building blocks:
- Mixture-of-Value Attention (MoVA): Attention is the part of a model that decides which earlier words matter for the next one. MoVA applies the expert principle to one part of it, the so-called value projections. The goal is better reasoning without extra compute.
- Softplus attention gating: A more flexible, theoretically grounded "dial" that controls how much information flows through attention.
- Uno: Lossless diffusion adapters that generate several tokens in parallel blocks instead of one at a time. According to the paper, this makes text generation up to 2.2x faster, and it doesn't need a separate helper model (draft model) the way standard speculative decoding does.
The speed claims don't match: the press release cites roughly 3x faster generation for "diffusion distillation," while the paper says "up to 2.2x." The paper's figure is the more conservative one and probably the more reliable.
The IFM uses the 36B-A4B model to show what MoVA and gating deliver. By its own account, the model nearly matches a dense 32B model trained on the same data, while using only a seventh of the active parameters.
What "fully open" means here
The IFM follows the open-source definition of LLM360, an initiative for transparent language models. The team released earlier K2 models the same way. This time, the package includes:
- checkpoints from every training phase
- pretraining data and training logs
- all training, evaluation and infrastructure code, including the xLLM framework for training dense and MoE models at production scale, and RL360 for reinforcement learning, which trains a model by rewarding good answers
- every pipeline used to build the training data
That last point is rare. According to the paper, it's the part of model development that open releases disclose least often. These pipelines produced around 10 trillion synthetic tokens, meaning training text generated by other models. The IFM says this shows synthetic data can make up a substantial share of pretraining at scale. Four of the six models were trained on the same 22-trillion-token corpus with the same recipe. That lets researchers cleanly study the effects of model size and sparsity (how much of the network is active at any given time).
IFM founder Eric Xing puts it this way:
"Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result and improve on it."
Where the openness hits its limits
The tech site shattered.io points out some weak spots. Citing a technical review, it reports that the 0.9B model's metadata at one point carried an "internal-only" license tag, and a LICENSE file was missing entirely. That suggests the IFM hasn't fully aligned its openness pitch with the actual license terms yet. Weights and code are under Apache 2.0, but the datasets each have their own licenses. Not everything the models were trained on can be freely redistributed. According to shattered.io, the release includes training data only "where the license allows." So if you want to rebuild a model completely from scratch, you'll have to work through a patchwork of different licenses.
Availability and support
The models are on Hugging Face and work from day one with the popular inference engines vLLM and SGLang, tools for running language models efficiently. According to shattered.io, there's also support from Ollama (for the 0.9B model) and Unsloth (for the 7B model). If you'd rather not self-host, you can use K2 Horizon via API through partners such as Compass, Cerebras, AWS and Nebius.
What you need to try it
The hardware you need depends heavily on the model. The sources don't give official memory requirements. The figures below are rough estimates for 4-bit quantization, which shrinks a model by storing its numbers at lower precision:
- 0.9B and 3.7B: run on almost any computer, even on the CPU alone or on an Apple Silicon Mac. About 1 to 4 GB of memory is enough.
- 7B: a graphics card with about 6 to 8 GB of VRAM, or a Mac with 16 GB of RAM. A free Google Colab GPU should also work.
- 32B and 36B-A4B: about 20 to 24 GB of VRAM, so an RTX 3090 or 4090, or a Mac with 32 GB or more. The 36B-A4B is much faster per token but still needs room for all its parameters.
- 375B-A23B: belongs in a data center, on multiple large GPUs. For personal use, the API is the realistic option.
Depending on the model, plan for anywhere from about 1 GB to well over 100 GB of free disk space for the download. Longer context, meaning more text in the model's "memory," takes additional memory.
The bigger picture
K2 Horizon faces strong competition: Alibaba's Qwen, Mistral and DeepSeek also release capable models under permissive licenses. According to shattered.io, that's why the IFM isn't claiming to win every benchmark. Its pitch is reproducibility. That's valuable for research: if you want to verify benchmark results, test safety measures or study how training data shapes a model, you need the data and the checkpoints, not just the finished model. Regulators and companies in regulated industries also increasingly want to know what a model was trained on.
Some questions remain open. The "best model in its size class" claims are vendor statements so far and need independent testing. The speed figures contradict each other (2.2x in the paper, about 3x in the press release). There were licensing inconsistencies at launch. And the label "largest fully open release in AI history" comes from the press release, not a neutral source. Still, K2 Horizon is one of the most comprehensive open model releases to date. It puts pressure on other labs that publish weights but keep their data pipelines secret.
Sources
- Hugging Face Papers huggingface.co
- finance.yahoo.com finance.yahoo.com
- shattered.io shattered.io
- prnewswire.com prnewswire.com
Try it yourself: K2-Horizon-7B
Explained step by step, no prior knowledge needed – on your own computer or free in your browser.
What can I do with it?
K2 Horizon is a family of text models built mainly for reasoning, math, coding and agent tasks, meaning tasks where the model plans several steps and uses tools. Running locally, you can chat with it, summarize text, have it write and explain code, or work through math problems step by step. According to the IFM, the 7B model is the best model under 10 billion parameters and is especially strong at software development and research. There are no independent tests of this yet. The small models (0.9B, 3.7B) are designed for low-end devices and, as you'd expect, are weaker at general knowledge and long texts. Because all the training data is open, the models are particularly well suited to research and your own fine-tuning. According to the sources, they don't handle images or audio, only text.
Who is it worth it for?
K2 Horizon is interesting for developers, researchers and tinkerers who want a capable model with traceable origins and a license that allows commercial use, and for anyone who wants to fine-tune their own model. If you just want the most convenient everyday chatbot, established models like Qwen in Ollama or LM Studio are the easier choice for now, because tool support there is more mature.
Ideas: What you can do with it
Run the 7B or 36B-A4B model on your machine and plug it into your code editor. It explains unfamiliar code, writes tests and finds bugs. Your source code never leaves your computer.
Enter homework or practice problems and have it walk you through the solution step by step. That way you see where you get stuck instead of just the final answer.
Use the 3.7B model and tools like Unsloth to fine-tune it on your own texts, such as your club's FAQ or your technical documents. You end up with a specialized assistant under the Apache 2.0 license.
The 0.9B model also runs on old laptops or single-board computers like a Raspberry Pi. For simple summaries, rephrasing or quick questions, it works with no internet connection at all.
Since the data and checkpoints are open, you can compare how the model changes between training phases. That's handy for a thesis or for experiments with synthetic data.
What your computer needs
Which variant is right for me?
Many models come in several sizes or "quantizations." These are compressed versions: smaller and faster, but slightly less accurate. Rule of thumb: pick the largest variant whose memory requirement fits in your VRAM (video memory). The scripts below choose this for you automatically.
| Variant | Download | Needs approx. | Compatible graphics cards |
|---|---|---|---|
Q4_K_M Default |
5.6 GB | 7.5 GB | e.g. RTX 3060 Ti, RTX 4060, RTX 5060, RX 7600 |
Q5_K_M |
6.5 GB | 8.5 GB | e.g. RTX 3060 12 GB, RTX 4070, RTX 5070, RX 7700 XT |
Q6_K |
7.4 GB | 9.5 GB | e.g. RTX 3060 12 GB, RTX 4070, RTX 5070, RX 7700 XT |
Q8_0 |
9.6 GB | 12 GB | e.g. RTX 3060 12 GB, RTX 4070, RTX 5070, RX 7700 XT |
BF16 |
18 GB | 21 GB | e.g. RTX 3090, RTX 4090, Tesla P40 |
For regular PCs and laptops, the 7B model is the best place to start (about 6–8 GB of VRAM at 4-bit, or a Mac with 16 GB). For very weak machines or CPU-only use, go with 0.9B or 3.7B. With a 24 GB graphics card (RTX 3090/4090) or a Mac with 32 GB or more, you can try 32B or 36B-A4B. The 36B-A4B responds faster because only 4 billion parameters do the work per token. The 375B-A23B is best used via the partners' APIs (e.g., Cerebras, AWS, Nebius). All VRAM figures are estimates, not official specs.
- 7B at 4-bit: roughly 6–8 GB of VRAM. At full bfloat16 precision (as in the vLLM example), it's about 15 GB or more.
- 0.9B and 3.7B also run without a graphics card, on the CPU or on Apple Silicon Macs.
- 32B and 36B-A4B need about 20–24 GB of VRAM at 4-bit. Even though the MoE model has only 4B active parameters, all 36B have to fit in memory.
- The 375B-A23B model requires data center hardware with multiple large GPUs.
- Long context takes extra memory. If you run out, reduce the context length.
How do I find out what my computer has?
- Windows: Press Ctrl + Shift + Esc → Task Manager → Performance tab → click GPU on the left. At the bottom you'll see Dedicated GPU memory (that's your VRAM). You'll find system memory under Memory.
- Mac: Apple menu at the top left → About This Mac. You'll see your chip (e.g. M2) and memory (e.g. 16 GB). On Apple chips, graphics and processor share this memory.
- Linux: Type
nvidia-smiin the terminal (shows your graphics card and its memory) andfree -hfor system memory.
Guide
Google Colab (in your browser, no installation)
Google Colab is a free service from Google. It gives you a computer with a graphics card in your browser for a few hours. All you need is a Google account. For this model: runs for free – The free Tesla T4 GPU with 15 GB is enough.
- Download the notebook. Click the button and save the file.
- Open Colab. Go to colab.research.google.com and sign in with your Google account.
- Upload the notebook. In the window that appears, click Upload on the left (or choose File → Upload notebook at the top) and select the file
colab.ipynb. - Turn on the GPU. In the top menu, go to Runtime → Change runtime type → under Hardware accelerator, select T4 GPU → Save.
- Start everything. Runtime → Run all (keyboard shortcut Ctrl + F9). If you see the warning “This notebook was not authored by Google,” click Run anyway.
- Wait. The cells run one after another. A green checkmark appears to the left of each cell when it's done. Installation and download take about 6 minutes here.
- Try it out. The result appears below the last cell. To test something else, change the text in the cell and run it again with Shift + Enter. The last cell starts a proper chat with an input field.
- Clean up. When you're done: Runtime → Disconnect and delete runtime. This saves your free quota.
Good to know: The free GPU is limited (usually a few hours per day). After about 90 minutes of inactivity, Colab disconnects, and you'll have to start everything over.
Windows 10 / 11
The easiest way is Ollama, a free program that launches AI models with a single command.
- Update your graphics driver (NVIDIA cards only): Install the NVIDIA app or the latest driver. For AMD cards, use the latest Adrenalin driver.
- Install Ollama. Download the file OllamaSetup.exe from ollama.com/download/windows, double-click it, and choose Install. Afterward, a small llama icon appears in the taskbar at the bottom right.
- Open a terminal window. Press the Windows key, type PowerShell and press Enter. A blue or black window opens.
- Start the model. Copy this command ("Copy" button), paste it into the window with a right-click, and press Enter:
The first time, Ollama downloads the model (5.6 GB (about 7 min at 100 Mbit/s)). After that, it's instant.ollama run hf.co/IFM/K2-Horizon-7B-GGUF:Q4_K_M - Start typing. Once
>>> Send a messageappears, type your question and press Enter. Type/byeto end the chat.
Not enough memory? Pick a smaller variant from the table above, e.g. ollama run hf.co/IFM/K2-Horizon-7B-GGUF:Q4_K_M.
Prefer automatic? A script that picks the right variant for you
- Download both files into the same folder: try_windows.ps1 demo.py
- Open the folder in File Explorer, right-click an empty spot and choose Open in Terminal.
- Paste this command and press Enter:
powershell -ExecutionPolicy Bypass -File try_windows.ps1
If needed, the script installs Ollama via winget, measures your VRAM (video memory) and downloads the matching variant.
macOS
Macs with an Apple chip (M1 to M5) are ideal. Older Intel Macs are very slow.
- Install Ollama. Download it from ollama.com/download/mac, open the file, and drag Ollama into your Applications folder. Launch Ollama once – a llama appears in the menu bar at the top.
- Open Terminal. Press Cmd + Space, type Terminal, and press Enter.
- Start the model. Paste the command (Cmd + V) and press Enter:
The first time, the model is downloaded (5.6 GB (about 7 min at 100 Mbit/s)).ollama run hf.co/IFM/K2-Horizon-7B-GGUF:Q4_K_M - Start typing. Type your question and press Enter. Quit with
/bye.
Linux
- Check your GPU.
nvidia-smishould show your card. If it doesn't, install the driver on Ubuntu and restart:sudo ubuntu-drivers install - Install Ollama:
curl -fsSL https://ollama.com/install.sh | sh - Start the model:
Exit withollama run hf.co/IFM/K2-Horizon-7B-GGUF:Q4_K_M/bye. Useollama psto check whether the model is running on the GPU. - Or let it pick the right variant automatically:
bash try_linux.sh # mit Python/transformers statt Ollama: MODE=hf bash try_linux.sh
Try it
Explain in three sentences what a Mixture-of-Experts model is.A train leaves at 9:40 a.m. at 120 km/h, and a second train leaves at 10:10 a.m. at 160 km/h on the same track. When does the second train catch up with the first? Work it out step by step.Write a Python function that finds all duplicate files in a folder based on their hash, and explain the code.Here's an error message from my program: [paste error message]. What's the likely cause, and how do I fix it?Plan a research project on whether a heat pump is worth it in an old house built in 1960: what sub-questions, sources and calculations do I need?
Tips for better results
- Start with the 7B model and only move up to 32B or 36B-A4B if the answers aren't good enough. The family is explicitly built for that path.
- For math and logic problems, explicitly ask for a step-by-step solution. That's what the models were trained for.
- For your own apps, you can launch the model with vLLM as an OpenAI-compatible API (e.g., on port 8000). Many existing tools then work right away.
- If you want to fine-tune, the IFM says the 3.7B model is especially well suited. It trains quickly and needs little hardware.
If something goes wrong
New architecture (MoVA, softplus gating): older versions of vLLM, SGLang or Transformers may not recognize the model. Always install the latest version.
According to reports, only the 0.9B model was available for Ollama at launch. For the larger variants, you may have to wait for community GGUF files.
The license information was inconsistent at launch. The 0.9B model was missing its LICENSE file for a while. Check the Model Card before any commercial use.
Weights and code are under Apache 2.0, but the training datasets have their own licenses. If you want to reuse the data, you'll need to check each one individually.
The Uno speed feature (diffusion adapters) may not be supported by every tool. Without Uno, the model runs at normal speed.
“out of memory,” a crash, or your computer freezes
The model is too big for your VRAM (video memory). Pick a smaller variant from the table above and close other programs, especially games and browsers with lots of tabs.
It runs, but extremely slowly
Your CPU is probably doing the work instead of your graphics card. In Ollama, the PROCESSOR column of ollama ps shows how much runs on the GPU. If it says "100% CPU", update your graphics driver or pick a smaller variant that fits entirely into VRAM (video memory).
“ollama” or “python” is not found
Close the terminal window and open it again so it recognizes the newly installed program. For Python on Windows: reinstall it and check Add python.exe to PATH.
Windows: "Running scripts is disabled on this system"
Run the script with exactly the command above: powershell -ExecutionPolicy Bypass -File try_windows.ps1. This allows execution for this one time only.
Error message when loading the model (version, "trust_remote_code", unknown architecture)
Brand-new models often need the latest software version. With Ollama: update Ollama. With Python: pip install -U transformers in the activated environment. The Model Card usually lists the exact version.
Tokenstrom generates the scripts and notebook automatically. Please take a quick look before running them – they're plain, readable text. Mind the model's license: apache-2.0.
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.