Kanana Explained: Everything About Kakao's AI — Models, KakaoTalk Agent & More 🐣

1. So What Exactly Is Kanana? 🤔

The name "Kanana" is a portmanteau of Kakao + Native + Natural, meant to evoke "the AI that's most like you." At first glance it might sound like just another chatbot name, but the reality is bigger than that. Kanana functions as a brand umbrella that ties together Kakao's AI efforts across the board.

Within the service, Kanana splits into two personas. "nana" operates in one-on-one chats, remembering personal context just for you, while "kana" lives in group chats, remembering only what happened in that specific room. For example, nana might recall a conference schedule and packing list discussed in a group chat and later remind you privately, while kana might quiz a study group on a paper they read together — different roles for different contexts.

The timeline moved fast. A little over a year after the standalone "Kanana" app quietly began private testing, the technology found its way into KakaoTalk itself and even replaced "hash search," a feature that had been around for more than a decade.

There's a clear business logic behind Kakao's all-in push. KakaoTalk already commands close to 50 million monthly active users, but the company has long faced criticism for relying too heavily on ads and commerce for monetization. At the same time, global tech giants like OpenAI and Google are rapidly expanding their own footprint with Korean users — and if Kakao fails to turn its messenger into an AI super-app, it risks losing control of the conversation itself, literally and figuratively. Seen this way, Kanana isn't just a new business line; many analysts view it as both a defensive and offensive move to protect KakaoTalk, Kakao's most valuable asset.

2. The Kanana App vs. "Kanana in KakaoTalk": What's the Difference? 🧭

Here's where a lot of people get confused: is the AI inside KakaoTalk the same thing as the separate app you download called "Kanana"? Short answer: they share a name but are two different products.

The standalone Kanana app is an independent, messenger-like service where you chat with AI companions — group mate kana and personal mate nana — much like talking to a friend. On first launch, the flow goes: set up your profile → connect a Kanana ID linked to your Kakao account → name and customize your own AI character → done. Reviewers have noted this onboarding is a fairly deliberate attempt to build emotional closeness with the AI. An October 2025 update reorganized the home screen into a three-tab layout (Home, Discover, Settings) and introduced interest-based "special AI mates." By December 2025, the app added memory that personalizes conversations based on your chat history, plus an "AI Studio" feature for generating group photos. During the 2025 KBO postseason, Kakao even trialed "biased broadcast" group chats that cheered exclusively for one team — a hit with baseball fans.

"Kanana in KakaoTalk," on the other hand, requires no separate download at all — it's an on-device AI agent built directly into the KakaoTalk app you already use. Early beta access was limited to high-end phones like the iPhone 15 Pro and above, but Kakao has gradually widened device support as it optimizes the on-device model for lower memory use. In short: the Kanana app is a new space to hang out with AI, while "Kanana in KakaoTalk" simply makes the KakaoTalk you already use smarter.

3. The Kanana Model Family, Fully Explained 🧩

Behind the scenes, several language models divide up the work of powering Kanana. The names can blur together, so here's a table to keep things straight.

What stands out most is Kanana-2's efficiency. According to Kakao, the model was trained on roughly 40% of the data used for China's GLM-4.5-Air-Base from Zhipu AI, yet still scored comparably on general benchmarks like MMLU. The reasoning-focused variant, "kanana-2-30b-a3b-thinking," was also tuned to minimize the common trade-off where tool-calling ability degrades during chain-of-thought reasoning. Kakao moved fast here too — an improved version (2601) shipped less than a month after the initial release, and a 155B-A17B flagship model is reportedly already in training. That said, it's worth remembering these figures come from Kakao's own benchmarking.

Digging a bit deeper technically, Kakao redesigned its tokenizer from scratch to improve Korean-token processing efficiency by about 30%, and adopted a Multi-head Latent Attention (MLA) architecture so the model can respond quickly even under KakaoTalk's traffic spikes. Language support isn't limited to Korean and English either — the models cover six languages total, including Japanese, Chinese, Thai, and Vietnamese. A separately released vision (image-understanding) model reportedly outperformed similarly sized models like Phi-3, Qwen 2.5, and HyperCLOVA X 3B on benchmarks.

The multimodal model "Kanana-o" is arguably the most future-facing piece of the puzzle. Because it lets you speak instead of type, it's well suited to hands-free scenarios — controlling navigation by voice while driving, or having messages read aloud when your hands aren't free. Cutting response latency from 1.5 to 0.5 seconds was specifically aimed at these kinds of real-time voice interactions. Internally, Kakao has described this model as the "sensory organ" of its agentic AI strategy — the next step after a Kanana that has, until now, been mostly limited to text.

4. Living With "Kanana in KakaoTalk" 💬

The Kanana most people actually encounter lives right inside their KakaoTalk chat rooms — "Kanana in KakaoTalk." There's no separate app to open; it watches the conversation and speaks up when it thinks it can help. Here's a mock-up reconstruction of a typical exchange.

▲ A reconstructed example based on how the actual service flow works.

The core features behind this flow break down as follows.

• Chat summaries: Open a chat room with five or more unread messages and Kanana automatically distills the key points.
• Proactive messages: Even without being asked, Kanana notices schedule-related context and reaches out first with reminders.
• Place agent: Initially focused on restaurant recommendations and bookings, a late-June 2026 update expanded coverage to tourist attractions, lodging, exhibitions/shows/movies, gas stations, convenience stores, and auto repair shops — recommend, explore, and book, all in one conversation.
• Kanana Search: An AI search feature you can use without leaving the chat room, which since April 2026 has been gradually replacing "Daum hash search," a feature that had been around for over a decade.
• Kanana Templates: Turns photos shared in chat into short AI-generated videos — useful for restoring old photos or animating a pet picture.

Privacy is worth addressing directly. Kakao's official pages state that raw chat text is neither transmitted to nor stored on its servers — conversation analysis is instead handled entirely on-device by the "Kanana Nano" model, according to the company. That said, because a messenger is such an inherently private space, some users remain uneasy about AI involvement in their conversations at all, regardless of where the processing happens.

Weighing feedback from users who've lived with the feature for a few weeks, the pros and cons roughly break down like this.

• 👍 Pros: No need to open a separate app — it blends naturally into the KakaoTalk you already use. In group chats, everyone gets the same summarized plan instead of everyone searching separately.
• 👍 Pros: Expanded place recommendations (lodging, exhibitions, gas stations) mean less manual work when planning a trip.
• 👎 Cons: Proactive messages occasionally misread context, coming across as an unnecessary notification.
• 👎 Cons: On some older devices, on-device processing means slower responses or limited functionality.

Device support keeps expanding. Early betas were limited to flagship phones, but Kakao has steadily widened compatibility as it shrinks the on-device model further. "Kanana Search" followed a similar path — starting with a small opt-in beta before gradually rolling out wider — and reviewers note that because it's built on Daum's huge content database, it feels meaningfully different from a typical web search.

5. ChatGPT for Kakao vs. Kanana: The Two-Track Strategy 🎭

Open KakaoTalk and you'll notice Kanana isn't the only AI in there — ChatGPT is too. In October 2025, Kakao partnered with OpenAI to launch "ChatGPT for Kakao." A single button at the top of the chat tab summons GPT-5-powered ChatGPT, which connects to Kakao's own services — KakaoMap, KakaoPay, KakaoT, Melon, gifting, and reservations — through a feature called "Kakao Tools." Ask something like "find me a croffle place near Hapjeong Station" and it will automatically pull up KakaoMap results complete with location data.

In February 2026, a limited-time promotion let users gift a ChatGPT Pro subscription — normally worth roughly 300,000 won — for just 29,900 won, a tenth of the price. The deal spread quickly through online communities, with plenty of "I bought it, no question" posts. Subscriber numbers reflect the momentum: ChatGPT for Kakao grew from 2 million users in November 2025 to 11 million by Q1 2026.

This growth is showing up in the numbers too. In its Q1 2026 earnings call, Kakao credited expanding AI services — including both Kanana and ChatGPT for Kakao — with contributing to double-digit year-over-year operating profit growth, particularly in its Talk Biz division. Internally, the company has also created a dedicated AI organization and kept hiring to support both tracks simultaneously.

6. Kanana vs. HyperCLOVA X: Korea's AI Rivalry 🇰🇷

Zoom in on the domestic market and Kakao's biggest rival is Naver. Naver released Korea's first (and the world's third) self-developed large language model, HyperCLOVA, back in 2021 and has steadily refined it since. Kakao, by comparison, is a relative latecomer with Kanana. The strategic differences between the two companies are pretty stark.

Early open-source adoption tells an interesting story. As of May 2025, cumulative Hugging Face downloads for the three Kanana Nano 2.1B variants sat at around 15,000, while Naver's HyperCLOVA X Seed 3B was closing in on 100,000 downloads in the same period. Naver's earlier move to release its model under a fully commercial-friendly license appears to have won over more developer attention early on. Kakao has since worked to close that gap by releasing Kanana-2 with a commercial license and fine-tunable weights.

The two companies' differing fortunes also showed up in a government-led "national representative AI" selection program. According to industry press coverage, NAVER Cloud was selected after impressing evaluators with HyperCLOVA X's track record of real-world deployment and its academia-industry partnerships, while Kakao reportedly fell short after failing to present concrete enough evidence of Kanana's technical strengths. That's just the outcome of one evaluation program, though, and shouldn't be read as a final verdict on Kanana's overall competitiveness.

Throw LG AI Research's "EXAONE" series into the mix and Korea's domestic AI landscape starts to look like a genuine three-way race. EXAONE leans toward industrial applications like manufacturing and biotech, a different focus from Kakao's messenger-first approach and Naver's search-and-commerce orientation. Some industry observers argue that all three Korean players are, in practice, competing less on raw foundation-model capability and more on how well they can weave AI into their existing services.

Zooming out further, there's still a significant scale gap between global giants like OpenAI, Google, and Anthropic, and domestic players like Kakao and Naver. That's why both Korean companies have largely chosen a pragmatic path — partnering with global models while focusing their own efforts on Korean-language and in-house-service optimization — rather than trying to build everything themselves from the ground up.

7. What Real Users Are Saying 📝

Specs and strategy only tell half the story. Here's what actual users and online communities have said about living with Kanana day to day — and it's not all praise.

Industry commentary has been sharper at times. One opinion piece from a current-affairs outlet relayed market skepticism that Kakao has fallen behind both global tech giants and domestic rivals on core AI technology, describing Kanana critically as "little more than bolting external AI onto existing services." That's just one perspective, though — the rapid subscriber growth of ChatGPT for Kakao and Kanana's roughly 70% positive-reaction rate can just as easily be read as signs the market response isn't bad at all. Whether this amounts to "repackaging" or "practical orchestration" remains an open debate.

Taken together, the overall picture is a mix of curious approval — "this is genuinely convenient" — and lukewarm skepticism — "search or a notes app already does this fine." One consistent theme, though: satisfaction tends to run higher in group chats than in one-on-one conversations. Most people seem to agree that pulling scattered information into one place is a genuinely useful role for the AI to play.

8. For Developers: Trying the Open-Source Models 👩‍💻

Both the Kanana Nano and Kanana-2 series are open-sourced on Hugging Face, so developers can download and fine-tune them directly. Kakao says it released training-stage weights as well, allowing teams to adapt the models freely using their own data. A minimal local inference flow looks roughly like this.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "kakaocorp/kanana-2-instruct"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Recommend three Korean BBQ places in Hongdae for a group of four"}
]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

output = model.generate(inputs, max_new_tokens=300)
print(tokenizer.decode(output[0], skip_special_tokens=True))

One developer's write-up after testing Kanana Nano noted that some benchmark scores improved over the previous version while others actually dropped, making the results "not exactly straightforward to interpret." Early releases also lacked quick-start example code, forcing developers to adapt existing snippets themselves — a sign the open-source ecosystem is still maturing. For solo developers and startups, the fact that the MoE-based Kanana-2 has been optimized to run on widely available Nvidia A100-class GPUs is a particularly welcome detail.

On the licensing side, Kakao says both the Kanana-2 and Nano series weights are released under terms that allow commercial fine-tuning and redistribution. That's already led developer communities to experiment with small-business chatbots, internal document-summarization tools, and offline kiosk assistants that need to run without an internet connection — exactly the kind of lightweight application Kanana Nano is built for. Developers who need content moderation can also pair their build with the separate "Kanana Safeguard" model for an extra layer of safety.

9. Where Kanana Goes From Here 🚀

Alongside the Kanana 1.5 release, Kakao already previewed its successor, "Kanana 2.5," aimed at being an even more capable agentic AI. It's reportedly being developed in three sizes — 30B-A3B, 8B, and 1.3B — with the 30B-A3B variant said to match 32.5B-class models while activating only 3B parameters. A from-scratch 150B-class flagship model is also in the works; in an earnings call, Kakao's CEO described it as targeting best-in-class performance among similarly sized domestic and global LLMs.

The Google partnership announced in February 2026 is another thread worth watching. If OpenAI is handling the B2C software side, discussions with Google reportedly center on Android OS optimization, on-device AI environments, and — further out — next-generation wearable devices. The idea of KakaoTalk expanding beyond smartphones into something like AI glasses is slowly starting to feel less speculative.

On the multimodal front, expect wider availability of "Kanana-o." The model ran a three-month closed beta starting in late February 2026 before gradually being demoed to the public at offline events like the World IT Show. The image-generation model "Kanana Kollage" is following a similar path, building public awareness through hands-on booth experiences first. For developers, API access and waitlist sign-ups are opening in stages too, which should mean more third-party Kanana-powered services in the near future.

5 Practical Tips for Getting the Most Out of Kanana 💡

A few tips worth knowing that make a real difference in daily use.

1. Ask for a summary before scrolling: If you open a group chat after a few days away and see a triple-digit unread count, ask Kanana to summarize first instead of scrolling. It'll pull out key decisions and any remaining to-dos.
2. Be specific with place recommendations: Instead of "recommend a good restaurant," try "4 people, around ₩30,000 per person, needs a private room for a work dinner." Being specific about headcount, budget, and purpose noticeably improves recommendation accuracy.
3. Turn on calendar syncing for better proactive reminders: Several users report that schedule-related briefings are more accurate and better-timed when calendar integration is switched on.
4. Turn old family photos into short videos with Kanana Templates: Reviving an old photo shared in a family chat as a short animated video tends to get a great reaction — some have used it as a substitute holiday gift.
5. Hand off complex tasks to ChatGPT for Kakao: Kanana is strong at understanding chat context, but long-form writing or complex logical reasoning is still often more reliable with a general-purpose model like ChatGPT. Splitting tasks between the two based on their strengths is currently the most efficient way to use both.

Frequently Asked Questions ❓

Wrapping Up

Put it all together and Kanana isn't really a single "product" — it's closer to an entire AI ecosystem Kakao is building across chat, search, agents, and image generation. What stands out is how Kakao pairs a pragmatic two-track strategy with OpenAI and Google with its own on-device models built to protect KakaoTalk's home-turf data. Criticism that the underlying technology "still has a ways to go" and praise that it's "more useful than expected" both coexist right now — that's simply where Kanana stands today.

Ultimately, whether Kanana succeeds will likely come down to something less flashy than benchmark scores: how often users find themselves thinking, without hesitation, "I'll just ask Kanana" instead of opening a separate app. If you haven't tried the version that speaks up first inside your own KakaoTalk chats yet, why not say hello the next time you're trying to lock down weekend plans? 😊