Kakao’s Kanana Revolution: Open‑Source Multimodal and MoE Models Signal a New Era for AI

Artificial intelligence research has long oscillated between open collaboration and proprietary competition. Open‑source releases such as TensorFlow, PyTorch and Llama models have democratized access to powerful algorithms and accelerated innovation. Yet as generative AI becomes the engine of software platforms, the world’s most capable models remain largely locked behind corporate paywalls. Against this backdrop, a wave of open AI releases from South Korea has begun to challenge the status quo. On July 24, 2025, tech giant Kakao announced that it would open‑source two advanced language models: Kanana‑1.5‑v‑3b, a lightweight multimodal model that processes text and images, and Kanana‑1.5‑15.7b‑a3b, the country’s first Mixture‑of‑Experts (MoE) model. Both are available to developers via Hugging Face under an Apache 2.0 license, and both represent a bold bid to increase access to high‑performance AI while strengthening South Korea’s technological independence.

Introduction

Most AI headlines are dominated by U.S. and Chinese labs. Companies like OpenAI, Google, Anthropic and Alibaba are engaged in a closed‑door arms race to build ever‑larger models, but they rarely release the full weights to the public. Even open‑weight initiatives—where trained parameters are shared but training data and code remain proprietary—offer limited transparency. That is why Kakao’s decision to open its Kanana family, including a state‑of‑the‑art multimodal model and an MoE model with 15.7 billion parameters, stands out. It is not just a technical upgrade; according to Kim Byung‑hak, head of Kanana Alpha, the release is “a move toward deploying AI in actual services while advancing Korea’s AI independence”.

Kakao’s announcement reflects a broader national effort. South Korea has embarked on an Independent AI Foundation Model Project that encourages domestic firms to develop sovereign models rather than relying solely on foreign technology. In the past two weeks, SK Telecom (SKT) released AX 3.1 Lite, a 7‑billion‑parameter model optimized for mobile devices, and LG unveiled Exaone 4.0, a hybrid language‑and‑reasoning model. Together, these announcements signal a regional shift towards open ecosystems, cost‑efficient architectures and on‑device intelligence. This article dives deep into Kakao’s Kanana release, contextualizes it within the global AI landscape, and explores its real‑world applications and implications.

From Messaging App to AI Powerhouse

Kakao began as a mobile messaging company but has since evolved into a diversified digital conglomerate. Its flagship service, KakaoTalk, dominates South Korean messaging with more than 45 million users, and the company operates platforms spanning finance, entertainment and mobility. Recognizing the strategic importance of AI, Kakao established Kakao Brain in 2017 and has invested heavily in large‑scale models. The Kanana series, launched in 2024, was designed to be a family of scalable models tailored to Korean languages and culture. Earlier iterations included four open‑source models released in February and May 2025. With the July releases, Kakao has moved beyond text‑only models and embraced multimodality and Mixture‑of‑Experts architectures.

The timing is strategic. The White House’s AI Action Plan in the United States frames open‑source AI as a geopolitical lever, and China’s DeepSeek and Qwen models have raised the bar for open performance. South Korea’s government aims to ensure its companies remain competitive by supporting domestic model development and deployment. Kakao participates in this Independent AI Foundation Model Project, positioning its releases not just as corporate initiatives but as part of a national agenda. This context matters: an open, sovereign model ecosystem can reduce reliance on foreign cloud providers, protect data sovereignty, and create local jobs.

Decoding Kanana‑1.5‑v‑3b: Lightweight Multimodality with Muscle

Design and Parameterization

Kanana‑1.5‑v‑3b is a 3‑billion‑parameter multimodal model that can interpret both text and image inputs. While 3 billion parameters may seem small compared to behemoths like GPT‑4, the design prioritizes efficiency and targeted performance. It uses training techniques such as human preference alignment and knowledge distillation, in which smaller models learn from larger teacher models. These methods enable the compact model to achieve performance exceeding much larger peers on specific tasks. Kakao reports that Kanana‑1.5‑v‑3b outperforms domestic competitors on instruction‑following benchmarks by 128%, an impressive margin given its size.

Multimodal Capabilities

Most open‑source language models are text‑only; multimodal models that handle images and text simultaneously have been dominated by proprietary players. Kanana‑1.5‑v‑3b addresses this gap. Kakao notes that it recognizes cultural landmarks like Cheonggyecheon in Seoul, interprets data charts, and solves math problems. The model has been trained on Korean and English data and demonstrates strong visual understanding across both languages. According to Business Korea, its ability to understand documents expressed in images is comparable to OpenAI’s GPT‑4o. That level of performance, combined with an open license, makes the model valuable for researchers and local businesses seeking to build translation tools, document analyzers or educational applications.

Performance Benchmarks

Kakao evaluated Kanana‑1.5‑v‑3b against domestic and international models of similar size. In Korean benchmarks, it achieved the highest scores among public models. In instruction‑following tests—important for chatbots and digital assistants—it delivered 128% of the performance of comparable domestic models. Business Korea reports that Kanana‑1.5‑v‑3b’s Korean and English image comprehension rivals global multimodal models and that it performs similarly to overseas open‑source models on English benchmarks. These results suggest that careful alignment and distillation can close the gap between compact and large models without the heavy compute requirements.

Flexibility and Applications

Beyond benchmarks, Kanana‑1.5‑v‑3b is designed for flexible application. It can handle tasks such as image and text recognition, storytelling, identifying cultural landmarks, chart analysis and math problem‑solving. These capabilities mean the model can power educational tools, tourism apps, reading aids for visually impaired users, or creative writing assistants. Developers can fine‑tune the model on specialized datasets or integrate it with mobile applications thanks to its moderate memory footprint.interprets data charts**, and *solves math problems*. The model has been trained on Korean and English data and demonstrates strong visual understanding across both languages. According to Business Korea, its ability to understand documents expressed in images is *comparable to OpenAI’s GPT‑1o*. That level of performance, combined with an open license, makes the model valuable for researchers and local businesses seeking to build translation tools, document analyzers or educational applications.

Performance Benchmarks

Flexibility and Applications

The Mixture‑of‑Experts Leap: Kanana‑1.5‑15.7b‑a3b

What Is a Mixture‑of‑Experts Model?

Standard transformer models activate all parameters for every input. This makes them compute‑heavy and limits scaling. Mixture‑of‑Experts (MoE) architectures partition the network into multiple “experts” and use a routing mechanism to activate only a subset of experts for each input. This yields sparse inference, reducing compute while allowing the model to scale parameter counts. The concept was popularized by Google’s Switch Transformer and later adopted in models like DeepSeek V2 and Mistral’s mixture‑of‑experts variants. However, most MoE models remain proprietary or partially open.

Architecture and Upcycling

Kakao’s Kanana‑1.5‑15.7b‑a3b contains 15.7 billion total parameters, but only about 3 billion parameters are active during inference. The model uses an upcycling method: it converts existing multi‑layer perceptron layers into expert layers by replication. Essentially, the researchers took their 3 billion‑parameter model and expanded it into a larger MoE model without re‑training from scratch. This approach reduces training cost and leverages knowledge distillation from the smaller base model. The result is a model whose performance matches or exceeds the Kanana‑1.5‑8B dense model while consuming fewer compute resources during use.

Efficiency and Cost Reduction

By activating only a subset of experts for each input, Kanana‑1.5‑15.7b‑a3b offers cost‑efficient inference. Business Korea notes that it was designed for companies and researchers seeking high‑performance AI without the financial burden of running dense models. MoE models can scale parameter counts and capture diverse patterns—useful for tasks requiring specialized knowledge—while keeping inference cost manageable. Kakao expects this architecture to be particularly beneficial for commercial environments, where compute budgets are a constraint.

Practical Implications

In practice, the MoE model could enable sophisticated AI services at a fraction of the cost. For instance, a media company could deploy the model to summarize hundreds of news articles simultaneously without needing a dedicated GPU for each request. A startup building a customer service bot could leverage expert partitions tuned for different domains—billing, technical support, product recommendations—while serving them through a unified interface. Because the model remains open and is licensed under Apache 2.0, developers can fine‑tune experts on their own data and even commercialize the resulting systems.

Why Open Source and Open Licensing Matter

The difference between open‑source and open‑weight AI is often misunderstood. Open‑weight models provide access only to the trained parameters, while the code and training data remain closed; open‑source models release the code, weights and often the training methodology. Kakao’s Kanana models are released under the Apache 2.0 license, meaning developers can use, modify and distribute them—even for commercial purposes. This stands in contrast to models such as Meta’s Llama 3, which uses a community license restricting commercial use, or OpenAI’s planned open‑weight model that will not include training data.

Open licensing fosters innovation by allowing researchers to inspect architectures, reproduce experiments and build upon them. Startups can integrate Kanana models into products without worrying about royalty fees. Universities can study the models to teach students about multimodal and MoE architectures. In the long term, open releases challenge the market dominance of a few large labs and promote competition based on quality rather than mere access.

South Korea’s Independent AI Foundation Model Project: Towards Sovereign AI

Kakao’s open‑source move is part of a broader national strategy. South Korea’s Independent AI Foundation Model Project encourages domestic companies to build “sovereign” models that reduce reliance on foreign technology and data centers. The project provides funding and resources for training large models on Korean data and aims to address concerns about data privacy and geopolitical dependence. By open‑sourcing Kanana models, Kakao signals its commitment to this vision and invites the research community to contribute.

The project also recognizes that AI is more than technical prowess; it is a strategic asset. Sovereign models help countries maintain control over critical infrastructure, prevent supply chain disruptions and mitigate the risk that changes in licensing or export rules will affect local businesses. In a geopolitical climate where AI leadership is increasingly tied to national power, South Korea’s initiative exemplifies how smaller nations can shape the future by investing in open, community‑driven ecosystems rather than purely replicating the scale of U.S. or Chinese giants.

A Regional Wave of Open Innovation: SK Telecom and LG Join the Fray

Kakao’s announcement was not the only significant open‑source release from Korea this week. SK Telecom (SKT) unveiled AX 3.1 Lite, a 7 billion‑parameter model for on‑device use, and released it on Hugging Face. The model is a successor to AX 3.0 Lite, which powers SKT’s A.Dot voice assistant. Despite its compact size, AX 3.1 Lite performs comparably to larger models on Korean benchmarks and even outperforms them on cultural awareness tests. SKT trained the model from scratch using its TITAN supercomputing system, optimizing each component from the tokenizer to the inference layer to ensure data remains local. With 32 transformer layers, 32 attention heads and a context length of 32,768 tokens, the model offers a balance between speed and accuracy. Releasing it as open source enables developers to build on‑device AI assistants without sending data to the cloud, addressing privacy and latency concerns.

Meanwhile, LG AI Research introduced Exaone 4.0, described as South Korea’s first hybrid reasoning AI model. Exaone combines a large language model with a reasoning engine to deliver quick responses and step‑by‑step analytical reasoning. It excels at tasks such as solving math problems and has achieved top scores on South Korea’s CSAT math exam and the U.S. Mathematical Olympiad. LG is making Exaone 4.0 available through FriendliAI’s platform and is offering a lightweight, on‑device version alongside a full‑featured expert model. The company plans to provide the model free to educational institutions, encouraging adoption and experimentation.

These concurrent releases underscore a distinct trend: South Korea is investing in open, efficient AI models across multiple industries. By providing alternatives to proprietary systems, these firms aim to cultivate a local AI ecosystem that can develop independently while remaining globally competitive.

Global Context: Open‑Source vs. Proprietary AI

The debate over open source versus proprietary models is intensifying. Advocates argue that open models democratize access, enhance transparency and spur innovation. Critics worry about misuse, lack of safety controls and economic disruption for companies that invest billions in research. The Kanana releases highlight a middle ground: open models can be designed with clear, permissive licenses and still achieve performance comparable to proprietary systems. Through knowledge distillation and human preference alignment, Kakao has demonstrated that smaller, open models can outperform competitors in specific benchmarks.

Globally, open‑source models are catching up. Meta’s Llama family has evolved from open‑weight to more permissive licensing; Mistral’s models have gained traction in Europe; and China’s DeepSeek and Moonshot AI have released billion‑parameter MoE models. Recently, Alibaba announced Qwen3‑Coder, an open‑source coding model built on a Mixture‑of‑Experts architecture that outperforms domestic competitors and competes with U.S. models like Claude and GPT‑4. These developments suggest that open models are no longer second‑rate alternatives; they are competitive tools that challenge proprietary leaders. Kakao’s releases contribute to this momentum and may inspire other companies and governments to embrace openness.

Real World Use Cases

1. Education and Research

The Kanana models are well‑suited for educational applications. Kanana‑1.5‑v‑3b’s ability to analyze charts and solve math problems could power interactive learning apps. Teachers could use the multimodal model to convert diagrams into step‑by‑step explanations or to generate practice problems tailored to students’ skill levels. Universities can experiment with the open code to teach courses on deep learning, multimodality and sparse architectures without relying on closed‑source alternatives.

2. Tourism and Cultural Preservation

South Korea’s tourism industry could leverage the model’s capacity to recognize cultural landmarks. A mobile app could allow tourists to photograph historical sites and receive contextual stories, translations and travel tips in real time. The model’s multilingual capabilities ensure that foreign visitors can access local knowledge without language barriers. Because the weights are open, cultural organizations can fine‑tune the model on specific heritage datasets, ensuring accurate representation and reducing bias.

3. Creative Content Generation

Kakao highlights that Kanana‑1.5‑v‑3b can be used for fairy tale and poetry creation. Combining image recognition with narrative generation opens possibilities for interactive storytelling. For example, children could upload a drawing, and the model would generate a bedtime story featuring the depicted characters and scenes. Musicians and artists might feed the model visual mood boards and receive lyrical or descriptive text to accompany their work. Because the license is permissive, creators can incorporate the AI into commercial projects without restrictive fees.

4. Business Intelligence and Data Analysis

Charts and spreadsheets often contain critical information that is hard to parse quickly. Kanana‑1.5‑v‑3b’s chart comprehension ability can help analysts extract insights from graphs and presentations. A business intelligence tool could take screenshots of financial dashboards and generate written summaries or suggestions. The MoE model’s efficiency makes it practical for large‑scale inference across thousands of reports, enabling companies to deploy AI analytics internally without prohibitive cloud costs.

5. Mobile and Edge Computing

SKT’s AX 3.1 Lite demonstrates that open models can operate on mobile devices. By incorporating Kanana’s techniques—efficient parameterization and upcycling—developers could build on‑device chatbots that function offline, preserving user privacy and reducing latency. Voice assistants in automobiles, wearables or IoT devices could leverage such models to interpret voice commands, translate speech and provide contextual responses without sending data to remote servers.

6. Healthcare and Assistive Technology

Multimodal models can enhance accessibility. A Kanana‑based app could read medical instructions from images and explain them verbally in multiple languages, helping patients with visual impairments or limited literacy. In the hospital setting, clinicians could photograph lab results or handwritten notes and receive structured summaries. The model’s strong instruction‑following ability and fine‑tuning potential make it adaptable to regulatory requirements.

7. Chatbots and Customer Service

The instruction‑following strength of Kanana‑1.5‑v‑3b and the scalability of the MoE model can improve customer support bots. For companies that operate in multiple domains—banking, e‑commerce, social media—MoE experts could handle specialized queries, while the central router directs requests appropriately. Because these models support both Korean and English, businesses can serve domestic and international clients seamlessly.

Conclusion

Kakao’s decision to open‑source Kanana‑1.5‑v‑3b and Kanana‑1.5‑15.7b‑a3b is more than a product launch; it is a statement about the future of AI. By releasing models that combine multimodal understanding, instruction following and mixture‑of‑experts efficiency under a permissive license, Kakao demonstrates that cutting‑edge capability need not be locked behind corporate gates. The models achieved top domestic performance and challenge global leaders in specific tasks while remaining accessible to researchers, developers and entrepreneurs.

These releases also illustrate a broader movement within South Korea. The Independent AI Foundation Model Project seeks to build sovereign models to protect national data and compete globally. SKT’s AX 3.1 Lite and LG’s Exaone 4.0 show that open innovation can span mobile devices, reasoning engines and hybrid architectures. While U.S. and Chinese companies continue to dominate the headlines, the Korean wave of open models underscores that AI leadership can emerge from unexpected corners when openness and efficiency are embraced.

For businesses and developers worldwide, the Kanana models offer a glimpse into a future where powerful, flexible AI is available to anyone willing to build on it. They invite a rethinking of what it means to compete in the AI race: success may hinge not on hoarding capabilities but on sharing them and inspiring others to collaborate. As more organizations follow Kakao’s lead, the AI landscape may become richer, more diverse and more equitable.

FAQs

Q1: What are Kanana‑1.5‑v‑3b and Kanana‑1.5‑15.7b‑a3b?
Kanana‑1.5‑v‑3b is a 3‑billion‑parameter multimodal language model that interprets both text and images and excels at tasks such as instruction following, chart analysis and cultural landmark recognition. Kanana‑1.5‑15.7b‑a3b is a Mixture‑of‑Experts (MoE) model with 15.7 billion parameters, of which only about 3 billion are active during inference. It uses an upcycled architecture to provide high performance at lower compute cost.

Q2: How does the Mixture‑of‑Experts architecture improve efficiency?
MoE models divide the neural network into multiple “experts” and use a router to activate only a subset of them for each input. This sparsity reduces computational load while enabling larger parameter counts. In Kanana‑1.5‑15.7b‑a3b, activating only about 3 billion of its 15.7 billion parameters allows the model to perform comparably to larger dense models. It achieves cost‑efficient inference, making it practical for businesses and researchers.

Q3: How can developers access and use the Kanana models?
Both models are published on Hugging Face under the Apache 2.0 license, allowing anyone to download, modify and integrate them into applications. Developers can fine‑tune the models on their own datasets, deploy them on local servers or integrate them into cloud services. The open license permits commercial use without royalties.

Q4: What is the difference between open‑source and open‑weight models?
An open‑source model provides access to the code, weights and often training methodology under a permissive license. An open‑weight model shares the trained parameters but may restrict access to the code and data. Kakao’s Kanana models are fully open source, while other companies offer open‑weight models with more restrictive terms. The open‑source approach fosters transparency and community contribution.

Q5: What impact will these releases have on the AI ecosystem in South Korea?
Kakao’s open‑source models support the government’s goal of building a sovereign AI ecosystem. By providing high‑performing models tailored to Korean language and culture, they reduce reliance on foreign systems and encourage local innovation. Combined with SKT’s and LG’s open releases, these models create a foundation for startups, universities and enterprises to build AI services without barriers.

Q6: Do the Kanana models support languages other than Korean?
Yes. Kanana‑1.5‑v‑3b is trained on Korean and English and demonstrates strong performance in both languages. Its multimodal design allows cross‑lingual understanding of images and text. The model could potentially be extended to other languages through fine‑tuning or future releases.

Q7: What future developments can we expect from Kakao?
Kakao plans to release reasoning models in the second half of the year and to continue scaling the Kanana series. The company is exploring ultra‑large models at the global flagship level and aims to incorporate agentic capabilities for more autonomous AI systems. Developers can expect new models that combine reasoning and multimodality while maintaining openness.

Tags: open‑source AI, Kanana, Kakao, MoE models, multimodal AI, AI models, The Artificial