K2-Horizon-7B-Uno

We present K2-Horizon-7B-Uno, a diffusion-augmented LLM with two pathways in a unified architecture:

• The AR pathway uses the AR weights of the K2-Horizon-7B model.
• The diffusion pathway augments these weights with LoRA-based diffusion adapters.

This repository contains the adapter only. The base-model weights are hosted separately in IFM/K2-Horizon-7B.

Code

Selected evaluation scripts are available in scripts/k2_horizon. The full evaluation suite will be released soon.

Evaluation results

The main number is the benchmark score and the subscript is TPF. Bold marks the best score in each row. -- and N/A denote unavailable results.

Benchmark	Uno 8B	Nemotron 14B	Mercury 2 size N/A	Diff-Gemma 26B-A4B
Agentic Tool Use
τ3 Banking	25.8--	--	9N/A	--
τ2 Telecom	90.12.46	14.34.8	71N/A	68.118.8
τ2 Retail	67.13.00	5.63.1	--	65.523.7
Terminal-Bench v2.1	39.12.65	4.57.5	27N/A	14.714.1
Agentic Coding
SWE-bench Verified	70.13.08	0.81.5	--	18.75.6
Long-Context Reasoning
AA-LCR	68.02.85	7.31.1	36N/A	19.710.8
Science and Knowledge
Humanity's Last Exam	18.62.47	2.67.2	16N/A	9.214.1
GPQA-Diamond	77.12.71	40.47.6	77N/A	70.711.9
AA-Omniscience	14.32.23	11.011.1	20N/A	17.79.9
Math
GSM8K	95.42.88	93.16.1	--	95.128.9
MATH500	98.92.86	89.25.6	--	92.424.1
AIME-24	93.02.90	56.74.9	--	73.716.9
AIME-25	90.72.96	40.04.5	--	74.318.9
AIME-26	86.32.84	46.74.8	--	70.717.8
Coding
MBPP	84.12.44	73.85.3	--	80.115.5
HumanEval	95.22.66	84.87.5	--	95.128.2
Efficiency Summary
Average TPF	2.71	5.41	N/A	17.56
System Throughput	5,255	2,794	1,197	1,136
Per-request Throughput	405	290	769	836

Citation

If you find this model useful, please cite:

@misc{k2_horizon_7b_uno,
  title        = {K2-Horizon-7B-Uno},
  author       = {Institute of Foundation Models},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/IFM/K2-Horizon-7B-Uno}},
}