| Name | Name | Last commit date | ||
|---|---|---|---|---|
🤗 HF Models | 📕 Pre-Training Blog | 📕 Post-Training Blog | 📕 Teaser Blog
2026/01/15: 🤗 Released kanana-2-30b-a3b-2601 HF model weights.2026/01/15: 📕 Published blog posts (pre-training, post-training) about the development of Kanana-2 models.2025/12/19: 🤗 Released kanana-2-30b-a3b HF model weights and publised a teaser blog.Kanana-2, the latest open-source evolution of the Kanana model family, is designed specifically for Agentic AI, presenting substantial enhancements in tool calling, complex instruction following, and logical reasoning. This new version adopts a cutting-edge architecture featuring MLA (Multi-head Latent Attention) and MoE (Mixture of Experts). These innovations allow the model to utilize significantly fewer active parameters compared to the previous 32.5B model while delivering superior performance and ensuring high throughput. Furthermore, the model natively supports context lengths of up to 32,768 tokens, enabling it to maintain coherence when handling extensive documents or long-context interactions.
In addition, Kanana-2 now supports 6 languages, covering Korean, English, Japanese, Chinese, Thai, and Vietnamese. To support this expansion, Kanana-2 utilizes a newly trained tokenizer that demonstrates superior tokenization efficiency across these languages, including an improvement of over 30% specifically for Korean. Finally, to address advanced problem-solving needs, Kanana-2 introduces reasoning models capable of deliberate thinking and reasoning, achieving significantly enhanced performance in downstream tasks, especially when tackling hard problems.
Note
No Kakao user data was used for either pre-training or post-training.
kanana-2-30b-a3b series has the following features:
| Model | Download |
|---|---|
| kanana-2-30b-a3b-base-2601* | 🤗 HuggingFace |
| kanana-2-30b-a3b-mid-2601* | 🤗 HuggingFace |
| kanana-2-30b-a3b-instruct-2601 | 🤗 HuggingFace |
| kanana-2-30b-a3b-thinking-2601 | 🤗 HuggingFace |
| kanana-2-30b-a3b-base | 🤗 HuggingFace |
| kanana-2-30b-a3b-instruct | 🤗 HuggingFace |
| kanana-2-30b-a3b-thinking | 🤗 HuggingFace |
kanana-2-30b-a3b-base-2601 (prior to mid-training) checkpoint to contribute to the research community.kanana-2-30b-a3b-mid-2601 is identical to kanana-2-30b-a3b-base.
| Benchmark | Metric | Shot | kanana-2-30b-a3b-mid-2601 | kanana-2-30b-a3b-base-2601 | kanana-1.5-32.5b-base | Qwen3-30B-A3B-Base* |
|---|---|---|---|---|---|---|
| General Tasks | ||||||
| MMLU | acc | 5 | 75.44 | 74.83 | 76.76 | 81.14 |
| MMLU-Pro | acc | 5 | 56.14 | 52.61 | 52.40 | 61.83 |
| BBH | acc | 3 | 79.76 | 76.46 | 81.54 | 79.97 |
| SimpleQA† | acc | 5 | 29.70 | 29.13 | 26.95 | 26.47 |
| Mathematics Tasks | ||||||
| MATH | em | 4 | 54.40 | 48.86 | 47.68 | 62.58 |
| GSM8K | em | 8 | 82.71 | 76.57 | 85.14 | 88.10 |
| Coding Tasks | ||||||
| HumanEval | pass@1 | 0 | 75.29 | 71.34 | 75.59 | 53.32 |
| MBPP | pass@1 | 3 | 62.39 | 60.21 | 65.96 | 72.58 |
| Korean Tasks | ||||||
| KMMLU | acc | 5 | 62.15 | 61.98 | 61.56 | 62.25 |
| KoSimpleQA† | acc | 5 | 49.70 | 49.40 | 45.70 | 26.33 |
| HAE-RAE Bench (v1.0) | acc | 5 | 88.73 | 88.91 | 90.65 | 72.04 |
| MATH-Ko‡ | em | 4 | 54.07 | 45.58 | 47.42 | 58.20 |
| GSM8K-Ko‡ | em | 8 | 77.48 | 70.43 | 81.43 | 88.10 |
| MBPP-Ko§ | pass@1 | 3 | 61.55 | 57.29 | 65.41 | 66.84 |
| Long Context Tasks | ||||||
| RULER-4K | acc | 0 | 93.09 | 92.49 | 86.39 | 94.32 |
| RULER-8K | acc | 0 | 92.29 | 92.14 | 90.16 | 92.16 |
| RULER-16K | acc | 0 | 90.73 | 90.01 | 85.88 | 91.28 |
| RULER-32K | acc | 0 | 88.63 | 87.92 | 81.62 | 88.32 |
| Benchmark | Metric | kanana-2-30b-a3b-instruct-2601 | kanana-2-30b-a3b-instruct | kanana-1.5-32.5b-instruct | Qwen3-30B-A3B-Instruct-2507* | Qwen3-30B-A3B (non-thinking)* |
|---|---|---|---|---|---|---|
| Chat | ||||||
| MT-Bench | judge† | 8.30 | 8.42 | 8.23 | 8.71 | 8.38 |
| KoMT-Bench | judge† | 8.21 | 8.24 | 7.94 | 8.49 | 7.89 |
| Instruction Following | ||||||
| IFEval | prompt strict | 87.25 | 84.47 | 79.48 | 82.62 | 84.10 |
| IFBench | prompt strict | 48.30 | 41.84 | 38.78 | 30.27 | 29.25 |
| Multi-IF (EN) | acc | 77.88 | 75.81 | 68.51 | 77.93 | 81.03 |
| Multi-Challenge | acc | 35.16 | 34.80 | 19.05 | 41.76 | 27.84 |
| Tool Calling | ||||||
| BFCL-v3 (Live‡) |
pass@1 | 76.66 | 74.30 | 68.74 | 73.93 | 69.14 |
| BFCL-v3 (Multi-Turn‡) |
pass@1 | 38.63 | 35.38 | 11.38 | 38.77 | 11.88 |
| Code Generation | ||||||
| HumanEval+ | pass@1 | 81.10 | 79.88 | 79.88 | 86.59 | 87.20 |
| MBPP+ | pass@1 | 73.02 | 73.81 | 71.96 | 75.13 | 75.13 |
| Mathematics | ||||||
| GSM8K | em | 93.10 | 91.89 | 91.58 | 93.56 | 93.33 |
| MATH | acc | 88.56 | 86.26 | 77.92 | 90.96 | 87.20 |
| Reasoning & Knowledge | ||||||
| MMLU | em | 81.61 | 80.80 | 82.75 | 87.13 | 85.60 |
| KMMLU | em | 68.26 | 67.32 | 65.75 | 67.56 | 63.49 |
| GPQA Diamond | pass@1 | 52.53 | 42.93 | 42.42 | 54.55 | 50.51 |
| HAERAE-Bench (v1.0) | em | 75.57 | 75.57 | 65.34 | 53.41 | 57.39 |
gpt-4o-2024-08-06 as the judge model.Live denotes the average score of 6 live benchmarks, and Multi-Turn denotes the average score of 4 multi-turn benchmarks.
| Benchmark | Metric | kanana-2-30b-a3b-thinking-2601 | kanana-2-30b-a3b-thinking | Qwen3-30B-A3B-Thinking-2507* | Qwen3-30B-A3B (thinking)* |
|---|---|---|---|---|---|
| Reasoning & Knowledge | |||||
| MMLU-Pro | pass@1 | 74.2 | 75.3 | 80.8 | 78.5 |
| GPQA Diamond | pass@1 | 57.8 | 61.3 | 70.6 | 62.6 |
| Competition Math | |||||
| AIME 2025 | pass@1 | 74.0 | 72.7 | 82.3 | 70.7 |
| AIME 2024 | pass@1 | 79.0 | 78.3 | 91.0 | 82.7 |
| AIME 2024-Ko† | pass@1 | 75.0 | 25.3 | 80.3 | 72.3 |
| Code Generation | |||||
| LiveCodeBench | pass@1 | 58.8 | 60.8 | 68.3 | 62.3 |
| LiveCodeBench-Ko‡ | pass@1 | 51.2 | 9.4 | 66.3¶ | 61.5¶ |
| Instruction Following | |||||
| IFEval | prompt strict | 82.2 | 82.2 | 87.8 | 86.1 |
| IFBench | prompt strict | 47.8 | 42.3 | 47.6 | 36.7 |
| Tool Calling | |||||
| BFCL-v3 (Live§) |
pass@1 | 75.9 | 75.6 | 82.9 | 80.3 |
| BFCL-v3 (Multi-Turn§) |
pass@1 | 43.7 | 34.3 | 53.6 | 35.6 |
Live denotes the average score of 6 live benchmarks, and Multi-Turn denotes the average score of 4 multi-turn benchmarks.Note
For optimal results with the reasoning model, please adhere to the default parameters: temperature=0.6, top_p=0.95, top_k=20. We strongly advise against greedy decoding, as it may lead to performance degradation and infinite repetition loops.
vLLM is a fast and memory-optimized engine designed for high-performance LLM inference and serving.
For kanana-2-30b-a3b-instruct-2601,
vllm serve kakaocorp/kanana-2-30b-a3b-instruct-2601 --enable-auto-tool-choice --tool-call-parser hermesFor kanana-2-30b-a3b-thinking-2601,
vllm serve kakaocorp/kanana-2-30b-a3b-thinking-2601 --reasoning-parser deepseek_r1 --enable-auto-tool-choice --tool-call-parser hermesSGLang is a high-efficiency framework for serving LLMs and VLMs, enabling easy deployment of OpenAI-compatible API servers.
For kanana-2-30b-a3b-instruct-2601,
python3 -m sglang.launch_server --model-path kakaocorp/kanana-2-30b-a3b-instruct-2601 --tool-call-parser qwenFor kanana-2-30b-a3b-thinking-2601,
python3 -m sglang.launch_server --model-path kakaocorp/kanana-2-30b-a3b-thinking-2601 --reasoning-parser deepseek-r1 --tool-call-parser qwenCurrently, the config.json uploaded to HuggingFace is configured for token lengths of 32,768 or less. To process tokens beyond this length, YaRN must be applied. By updating the config.json with the following parameters, you can apply YaRN to handle token sequences up to 128K in length:
"rope_scaling": {
"beta_fast": 32,
"beta_slow": 1,
"factor": 4.0,
"mscale": 1.0,
"mscale_all_dim": 1.0,
"original_max_position_embeddings": 32768,
"type": "yarn",
},Passing command line arguments for deployment:
vllm
vllm serve ... --hf-overrides '{"max_position_embeddings": 131072, "rope_scaling": {"rope_type":"deepseek_yarn","factor":4.0,"beta_fast":32,"beta_slow":1,"mscale":1.0,"mscale_all_dim":1.0,"original_max_position_embeddings":32768}}'sglang
python3 -m sglang.launch_server ... --json-model-override-args '{"max_position_embeddings":131072, "rope_scaling":{"rope_type":"deepseek_yarn","factor":4.0,"beta_fast":32,"beta_slow":1,"mscale":1.0,"mscale_all_dim":1.0,"original_max_position_embeddings":32768}}'Note
Most leading open-source implementations of static YaRN apply a constant scaling factor, which can negatively impact performance on shorter texts. To ensure optimal performance:
rope_scaling only when necessary for processing long contexts.factor based on your specific needs (e.g., set factor to 2.0 for a 65,536-token context)."The model weights are released under the Kanana License.
@article{,
title={Kanana-2 LLM},
author={Kanana LLM},
year={2025},
url={https://huggingface.co/collections/kakaocorp/kanana-2}
}