K2-Horizon-0.9B-Uno

K2-Horizon-0.9B-Uno is a conditional-LoRA adapter for diffusion-style decoding with K2-Horizon-0.9B. This repository contains the adapter only. The base-model weights are hosted separately in IFM/K2-Horizon-0.9B.

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 accuracy and the subscript is TPF. -- denotes an unavailable TPF.

Benchmark	Uno 0.9B
Long-Context Reasoning
AA-LCR	18.01.81
Science and Knowledge
ARC-Challenge	78.21.46
GPQA-Diamond (avg@16)	27.31.59
HLE (Full)	5.41.53
AA-Omniscience	7.21.64
Math
AIME 2024	43.31.50
AIME 2025 (avg@16)	41.71.54
AIME 2026 (avg@16)	48.51.52
GSM8K	88.21.58
HMMT February 2026 (avg@16)	25.8--
MATH500 (Full)	86.21.57
Coding
HumanEval	62.81.79
HumanEval+ (pass@1)	79.9--
LiveCodeBench v6 (avg@3)	37.4--
MBPP	70.41.53
MBPP+ (pass@1)	68.0--
Instruction Following
IFEval (strict instruction)	80.8%1.72

Conversion provenance

conversion_summary.json records the adapter hash, source checkpoint, base-weight match, and tensor-key compatibility validation.

The public K2-Horizon base is a schema migration of the exact local base used for training. The base weight, index, and tokenizer hashes match; only the public Python architecture name changed from K2Aurora to K2Horizon. All 392 adapter tensors map to valid target weights in the public base model.

Citation

If you find this model useful, please cite:

@misc{k2_horizon_09b_uno,
  title        = {K2-Horizon-0.9B-Uno},
  author       = {Institute of Foundation Models},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/IFM/K2-Horizon-0.9B-Uno}},
}