Kakao Corp. published this model on January 15, 2026, with a native 32,768-token context that extends to 128K through a YaRN override. It suits teams that want a self-hosted Korean-and-English assistant with tool calling, and it is the wrong pick for anyone planning to resell access, which needs a separate license.
What changed
- · Release Kanana-2 instruct and thinking weights refreshed (2601 series) source
Kanana-2 30B-A3B Instruct is an open-weight mixture-of-experts chat model from Kakao, released in January 2026. It scores 48.30 on IFBench against 30.27 for Qwen3-30B-A3B-Instruct-2507 in Kakao's own table, activates 3B of its 30B parameters per token, and supports six languages including Korean.
Where it sits
- 52.5%% correctGPQA DiamondHigher is better#51 / 52peer median 87.6%per source, see benchmark scores
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Kakao · Family: Kanana
Context window: 32,768 tokens
Input modalities: text, tool-calls · Output: text, tool-calls
About Kanana-2 30B-A3B Instruct
Kanana-2 30B-A3B Instruct is the instruction-tuned chat model in the Kanana-2 family from Kakao, published on Hugging Face as kanana-2-30b-a3b-instruct-2601 with weights released on January 15, 2026. It follows the December 2025 instruct checkpoint and is described on its model card as designed for agentic work: tool calling, complex instruction following and logical reasoning. It is a mixture-of-experts Transformer with 48 layers, 128 routed experts (6 chosen per token) and 2 shared experts, using multi-head latent attention. Total size is 30B parameters with 3B active.
Kakao's own table compares the model with its predecessor Kanana 1.5 32.5B and with two Qwen3-30B-A3B variants. The strongest rows are instruction following: IFEval 87.25 and IFBench 48.30, against 82.62 and 30.27 for Qwen3-30B-A3B-Instruct-2507. On BFCL-v3 live tool calling it scores 76.66 against 73.93. It trails that Qwen model on MMLU (81.61 against 87.13), GPQA Diamond (52.53 against 54.55), MATH (88.56 against 90.96) and MT-Bench (8.30 against 8.71). On Korean tests it is ahead on HAE-RAE Bench (75.57 against 53.41) and narrowly on KMMLU (68.26 against 67.56). Every number comes from Kakao's internal evaluation toolkit and has not been independently reproduced by HokAI.
The native context length is 32,768 tokens. The card documents a YaRN setting that stretches it to 128K, with a warning that static scaling can hurt short-text quality, so it should be switched on only when needed. Supported languages are Korean, English, Japanese, Chinese, Thai and Vietnamese, and a new tokenizer is reported to improve Korean tokenization efficiency by over 30%. The model is text-in, text-out with tool calls; there is no image or audio input.
There is no hosted API price. The weights are free to download from Hugging Face, so the cost is hardware: at BF16 the 30.67B parameters come to roughly 61 GB before the KV cache, by simple arithmetic. The card gives vLLM and SGLang launch commands, including tool-call parsers, and the site lists the model as self-hosted only. The Kanana License allows free use for running your own service, but reselling access through an API, bundling it in on-premise or system-integration packages, or embedding it on-device needs a separate commercial license, as does any company over 10 million monthly active users.
The nearest comparisons are other open-weight models from Korean labs and mid-size mixture-of-experts models elsewhere: K-EXAONE 2.0 from LG AI Research, HyperCLOVA X from Naver, Solar Pro 4 from Upstage, and in size class Nemotron 3 Nano 30B-A3B and Gemma 4 31B. Browse Kakao's models or the open-weights group, or compare against the full model directory. The card says no Kakao user data was used for pre-training or post-training.
Pricing
Kakao publishes no per-token API price for this model on the pages HokAI checked. The cost is the GPU time to host the checkpoint; the first FAQ gives the weight size. Third-party hosts were not reviewed.
Key Features
- Mixture-of-experts design: 128 routed experts plus 2 shared experts per sparse layer, with 6 routed experts chosen for each token across 48 layers.
- Agentic tool calling: Scores 76.66 on BFCL-v3 live and ships vLLM and SGLang parser flags for automatic tool choice.
- Six-language support: Covers Korean, English, Japanese, Chinese, Thai and Vietnamese with a tokenizer Kakao reports as over 30% more efficient on Korean.
- Extendable context: Native 32,768-token window, extendable to 128K by applying the YaRN override in the model card.
Pros
- Open weights with a license that allows free use inside your own service, up to 10 million monthly active users.
- Leads Kakao's own table on IFEval, IFBench and BFCL-v3 live against Qwen3-30B-A3B-Instruct-2507.
- Only 3B active parameters per token keeps compute per request low for a 30B-class model.
Cons
- Trails Qwen3-30B-A3B-Instruct-2507 on MMLU, GPQA Diamond, MATH and MT-Bench in the vendor table.
- 32K native context, with 128K only through a configuration override.
- The Kanana License is not an OSI license; API resale, SI packaging and on-device embedding need Kakao's permission.
Benchmarks
- MATH: 88.6% vendor-reported · 13 Oct 2026 — Competition maths problems, % solved.
- MMLU: 81.6% vendor-reported · 13 Oct 2026 — General-knowledge exam across 57 subjects, % correct.
- IFBench: 48.3% vendor-reported · 13 Oct 2026 — How precisely the model follows detailed instructions, % passing.
- GPQA Diamond: 52.5% vendor-reported · 13 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What does it cost to run Kanana-2 30B-A3B Instruct?
Kakao publishes no per-token API price for it, so the cost is your own hardware. The checkpoint holds 30.67B parameters in BF16, which is roughly 61 GB of weights before the KV cache. The weights themselves are free to download from Hugging Face.
Does Kanana-2 beat Qwen3-30B-A3B on benchmarks?
It depends on the axis in Kakao's own table. Kanana-2 wins on tool calling (BFCL-v3 live) and Korean knowledge (HAE-RAE Bench), while the Qwen model wins on MMLU, GPQA Diamond, MATH and MT-Bench. Pick Kanana-2 for Korean agent work and Qwen for broad English knowledge.
Is Kanana-2 open source?
It is open-weight, not open source in the OSI sense. The Kanana License lets you use the weights in your own service for free, but selling access by API, shipping it in packaged on-premise deployments, or putting it on devices needs written permission from Kakao. Companies above 10 million monthly active users also need that permission.
Does Kakao train on my data if I use Kanana-2?
Self-hosting means your prompts never reach Kakao. Separately, the card says Kakao kept its own user data out of training. Kakao runs no hosted endpoint that HokAI could review, so there is no retention policy to report.
Who should pick Kanana-2 over a bigger model?
Teams with limited GPU capacity that need Korean, Japanese or English chat with tool calling should try it, since only 3B parameters are active per token. Teams that need long documents, images or the strongest math and knowledge scores should look at Qwen, Gemma or K-EXAONE instead.