K2-Horizon releases six open models with their full training recipe
A team at MBZUAI has published six language models for agent work, from phone-sized to server-sized, together with the data, code and checkpoints used to build them. The largest is competitive with open rivals on tool use, but trails closed models and has weak defences against hidden instructions.

AI illustrationWhat the study shows
The K2 Team at the Institute of Foundation Models, part of Mohamed bin Zayed University of Artificial Intelligence, released a family of six models called K2-Horizon. This is a technical report on arXiv. It is a preprint and has not been peer reviewed.
The models range from 0.9 billion parameters, small enough for a phone, to 375 billion. The largest is a mixture-of-experts model, where only part of the network runs for each word. It uses about 23 billion parameters per word. Alongside the final models, the team says it released 380 intermediate checkpoints, about 21.6 TB of training data and 17 code repositories.
The authors compared the largest model with four open models of similar vintage and three closed ones. It led the open models on Toolathlon-Verified, a test of multi-step tool use, scoring 65.3. On Terminal-Bench 2.1, where an agent works in a live command-line session, it scored 70.2 and placed second among open models. Its clearest weakness was expert-level science, where it trailed the best open model by about 9 points on the HLE exam.
A mid-sized model tests a new design called Mixture-of-Value Attention, which applies the expert idea to part of the attention mechanism. It uses about one seventh of the active parameters of a dense 32B model trained on the same data. The authors report it stays within a few points of that model on every benchmark they ran.
The team also built Uno, small add-on modules that guess several words at once while the base model checks them. The authors say this keeps output identical to the base model. For the 7B model it gave about 1.5 times more throughput when serving 64 requests at once, and about 2.2 times for a single request.
About 10 trillion of the training words were synthetic, generated by other models and mostly not checked for correctness. The authors argue that wrong answers in long worked examples are ordinary noise, and that strict filtering would cut diversity. They also chose their data mix by predicting results at full scale. That reversed the choice that looked best in small test runs.
Our reading
In our reading, the release itself is the main contribution. Benchmark wins are narrow and self-reported. But a full development history, from raw data to quantised builds, is rare at this scale. It gives researchers a way to study what each training step changes, which finished weights alone cannot do.
The report is also unusually frank about failures. One example: every coding trajectory placed the repository in the same folder. The model learned this habit so well that, when the folder moved, tasks took twice as many commands and the pass rate fell by 40%. Lessons like this are practical and transferable to any team building agent training data.
The performance picture is mixed. On agent tasks the largest model sits near the top of the open field with fewer active parameters than most rivals. On expert science and factual accuracy it is clearly behind. Against closed frontier models it trails on most rows of the authors' own table.
Read the rest of this story
Membership is free. You get every article in full and the Issue every Sunday.
We send a sign-in link, no password needed. Already a member? Sign inReadiness assessment
| Level | Status | Evidence |
|---|---|---|
| I Lab only | Met | The work goes beyond lab-only: models are trained at scale and results are reported. |
| II Benchmarked | Met | Results are reported on public benchmarks such as Terminal-Bench 2.1, Toolathlon-Verified, tau3-Banking, SWE-bench Verified, GPQA Diamond and HLE, with named open and closed baselines. They are run by the authors, not independently. |
| III Open source | Met | The paper says final weights, intermediate checkpoints, quantised builds, training data and 17 code repositories are released. Models and data are Apache 2.0 except TxT360-v2 (CC BY 4.0). Links given include huggingface.co/collections/IFM/k2-horizon and github.com/ifm-ai. I could not open these links. |
| IV In production | Not yet | The paper documents no named organisation using the models in production. It documents internal use only: the vendor pipeline built the models and ran evaluations. The authors' own deployment advice is conditional, with a human in the loop for agents. |
What would move it up
Documented production use by a named organisation, or independent third-party reproduction of the headline benchmark results. Level 3 also depends on checking that the released artefacts exist and are complete.
The study
K2 Team, Zhengzhong Liu, Haonan Li et al.. K2-Horizon: A Frontier Fleet of Open Source Language Models. arXiv, 16 November 2026. Read the paper
Notes
- K2 Team, Institute of Foundation Models. K2-Horizon: A Frontier Fleet of Open Source Language Models. arXiv preprint 2611.21003, 16 November 2026. Not peer reviewed. Source
- Project page listing all released artefacts. Source
- Model weights and checkpoints on Hugging Face (as stated in the paper; not yet checked by us). Source
- Code repositories, including xLLM, RL360 and Eval360 (as stated in the paper; not yet checked by us). Source
- Exact figures: 375B-A23B scores 65.3 on Toolathlon-Verified, 70.2 on Terminal-Bench 2.1, 48.4 on SWE-Atlas-QnA (strict), 34.0 on tau3-Banking, 32.0 on HLE (best open model 41.1). Reward-hacking audit: 24 of 500 passes flagged, pass rate 70.2% to 66.9%. Agentic safety: 2.5% full refusal under indirect prompt injection.
- Readiness: level III, open source available, pending confirmation that the listed artefacts are live.
Edited and signed by
Stefan PerdalEditor and publisher86 stories editedStefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.
How this was made
An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.
How we workNo corrections