Kanana

4 posts

kakao4 min readCurated summary

Smaller and More Powerful Kanana SLM Development

Kanana-2 is Kakao’s second Small Language Model series, designed to deliver strong performance under the memory and compute constraints of on-device environments. It includes 3B, 1.3B, and 0.9B models, using improved pre-training, distillation, pruning, tokenizer efficiency, and Sliding Window Attention. Kakao reports that the resulting models outperform earlier Kanana models and compare favorably with similarly sized open-source models. ## Motivation and Model Lineup - On-device services require models that are small and fast because smartphones have limited memory and compute. - Kakao uses its own SLMs in the “Kanana in KakaoTalk” service. - The Kanana-2 series consists of: - Kanana-2-3B - Kanana-2-1.3B - Kanana-2-0.9B - Base and Instruct versions of the 3B and 1.3B models are being released. - The development builds on techniques from Kanana-2-30B-A3B and earlier Kanana Nano models. ## Efficiency Improvements - **Kanana-2 Tokenizer** - Improves Korean tokenization efficiency by more than 30% compared with the previous tokenizer. - Reduces the number of tokens required for Korean text, improving processing efficiency. - **Sliding Window Attention** - Reduces KV cache size during inference. - Improves memory efficiency and helps reduce decoding bottlenecks on devices. ## 3B Pre-Training ### TPU-Based Training from Scratch - Kanana-2-3B-Base was initially trained from scratch on a TPU v5e cluster. - Kakao used a MaxText-based internal training framework. - The team developed infrastructure allowing training to transition between TPU and GPU clusters: - Pre-training was completed on TPU. - Distillation was subsequently performed on GPUs using Megatron-LM. - Pre-training used two stages: - Stage 1: 7.5 trillion tokens - Stage 2: 2 trillion tokens - The Muon optimizer was used throughout pre-training. ### Learning-Rate Scaling - Directly searching learning rates at multi-trillion-token scale would be too expensive. - Kakao instead tested learning-rate candidates using a 100-billion-token proxy dataset while preserving the Stage 1 data distribution. - The selected learning rate was scaled to the full 7.5-trillion-token training run using the Token Horizon scaling rule: `LR_target ≈ LR_proxy × (D_target / D_proxy)^−β` - The experiments used: - `D_proxy = 100B` - `D_target = 7.5T` - `β = 0.32` - This approach enabled stable hyperparameter selection with a smaller exploration budget. ## Teacher-Based Distillation - The Kanana-2-30B-A3B-Instruct-2601 model was used as the teacher. - Kakao compared Base, Instruct, and Thinking versions of the teacher model. - The Instruct teacher consistently produced the strongest student-model performance. - The result supports recent findings that post-trained teachers can be especially effective for transferring mathematical and coding capabilities. ## Long-Context Training - The model’s context length was expanded from 4K to 32K using YaRN. - Additional mid-training data was introduced during the learning-rate decay phase. - The resulting Kanana-2-3B-Base reportedly surpassed earlier Kanana 3B models across Korean and English knowledge, mathematics, and coding. - It also exceeded many similarly sized open-source SOTA base models. ## Building the 1.3B and 0.9B Models - Kanana-2-1.3B-Base and Kanana-2-0.9B-Base were progressively derived from Kanana-2-3B-Base. - The process extended the Minitron-based structured pruning and knowledge-distillation approach used for Kanana Nano. - The smaller models were trained with SWA-aware long-context procedures to preserve efficiency in on-device decoding. ### Improved Hidden-Dimension Pruning - Traditional hidden-dimension pruning scores each dimension independently using activation statistics from calibration data. - This is efficient but may overlook information represented jointly across multiple dimensions. - Kanana-2 applies PCA-based pruning inspired by Ministral 3: - Collect activation statistics from Attention RMSNorm, MLP RMSNorm, and Final RMSNorm inputs. - Compute a global rotation matrix with PCA. - Apply the rotation consistently to token embeddings and attention/MLP projection weights. - Reduce hidden dimensions after rotation, aiming to preserve more meaningful shared representations. Kakao’s approach combines large-scale pre-training, teacher distillation, structured compression, better Korean tokenization, and attention-level inference optimization. For practical on-device deployment, the smaller Kanana-2 models are intended to offer a better balance of speed, memory usage, and capability than simply deploying a larger general-purpose LLM.

Read original(opens in new tab)
kakao2 min readCurated summary

Recruiting new Kakao AI Ambassadors ‘

Kakao is recruiting 100 participants for its expanded KANANA 429 AI ambassador program. The five-month program introduces separate tracks for AI experts, creators, and university students, offering opportunities to test Kakao AI services, create content, and provide feedback. Applications close at noon on February 19, 2026. ## Program Purpose and Name - KANANA 429 promotes Kakao’s AI technologies and gathers user feedback. - “429” references the HTTP status code “Too Many Requests,” representing people with abundant enthusiasm and ideas about AI. - The program builds on Kakao’s first ambassador cohort, which included 20 participants. ## Results from the Previous Cohort - Participants communicated through KakaoTalk Open Chat. - They attended monthly offline meetups, informal group activities, and networking sessions with Kakao employees. - They previewed new Kakao AI services and exchanged feedback. - The cohort produced roughly 100 reviews and other pieces of content about Kakao’s AI services, models, and technologies. - Kakao selected and awarded five outstanding ambassadors. ## New Tracks and Benefits - **AI experts:** Test Kakao’s latest AI services and models and write in-depth reviews. - **Creators:** Produce content demonstrating practical ways to use Kakao AI. - **University students:** Promote the program on and off campus and collect user opinions. - The activity period has increased from three to five months. - Selected ambassadors receive AI service usage opportunities worth approximately 1 million won, along with additional benefits and special merchandise. ## Application and Schedule - Applicants must publish content related to Kakao AI and submit its URL through the recruitment page. - Applications are accepted until noon on February 19, 2026. - Every applicant receives a one-month free Kakao Emoticon Plus subscription. - Selected participants will be notified individually through Kakao’s official KakaoTalk channel on March 4. - The opening ceremony is scheduled for March 13 at Kakao AI Campus. Kakao is seeking applicants who are genuinely interested in AI and willing to communicate openly while helping shape and spread its AI services.

Read original(opens in new tab)
kakaoOriginal article

Kakao’s “ (opens in new tab)

Kakao's Kanana-v-4b-hybrid is a multimodal language model designed to transcend simple image-to-text conversion by integrating logical reasoning and self-verification directly into its response process. By employing a hybrid architecture that handles both intuitive dialogue and complex visual reasoning within a single model, it achieves high accuracy and reliability for sophisticated tasks. This approach allows the model to maintain consistency in user experience while excelling in Korean-specific contexts, as evidenced by its record-breaking 92.8 score on the KoNET evaluation. ### Integrated Hybrid Architecture * Consolidates intuitive tasks (like OCR and summarization) and logical tasks (complex reasoning) into a single model to reduce system complexity and maintenance costs. * Eliminates the need for external routing between specialized models, ensuring a consistent tone, response format, and safety policy throughout a single conversation session. * Utilizes a refined training recipe that balances data ratios and visual reasoning training to ensure that improvements in multimodal understanding benefit all types of user queries. ### Visual Reasoning and Self-Reflection * Follows a natural logic flow: synthesizing information from images and text, applying conditions, verifying candidates, and finally concluding the response. * Features a "Reflection" mechanism where the model actively monitors its own thought process to catch "small but fatal" errors, such as calculation mistakes or missed constraints. * Excels in high-stakes visual tasks like receipt auditing, table filtering, and mathematical problem-solving by double-checking intermediate results against original image data. ### Native Korean Logical Processing * Prioritizes "thinking in Korean" to accurately preserve the nuances of complex constraints, such as "except for X" or "only in cases of Y," which are often lost during internal translation. * Develops a native Korean Rationale process to prevent logical drift, ensuring that the internal reasoning steps remain perfectly aligned with the linguistic structure of the user's query. * Addresses the difficulty of processing information scattered throughout Korean-language documents or exam papers by synthesizing data without language-conversion overhead. Kanana-v-4b-hybrid marks a shift toward "verifiable AI" that provides evidence-based answers rather than just plausible text. For applications in education, finance, or complex document processing, this model offers a blueprint for building trust through transparent reasoning and self-correction.

kakaoOriginal article

The Evolution of Kanana-o Toward (opens in new tab)

Kakao has significantly advanced its integrated multimodal model, Kanana-o, by enhancing its ability to process complex instructions across text, image, and audio inputs while enriching its emotional vocal expression. By developing specialized datasets and sophisticated training techniques for prosody, the team has bridged the performance gap between text and audio modalities. The result is a more natural, human-like AI capable of nuanced interaction and high-performance instruction following, particularly within the Korean linguistic context. ## Advancing Multimodal Instruction Following * Addressed the "modality gap" where multimodal models often show decreased reasoning and reasoning performance when processing audio inputs compared to text. * Constructed a structured, high-quality dataset featuring complex, multi-step instructions such as summarizing a context and then translating it into a specific language or style. * Leveraged the Speech-KoMT-Bench to evaluate performance, showing that Kanana-o significantly outperforms global competitors of similar scale in Korean-specific tasks. * Focused on "Domain-generalization" to ensure the model's core intelligence remains stable regardless of whether the input is text, audio, or a combination of both. ## Image-Audio-Text Modality Alignment * Developed integrated datasets to ensure that reasoning capabilities learned in text-image or text-audio contexts generalize to complex image-audio scenarios. * Trained the model to handle tasks where users ask questions about visual information via voice, requiring the simultaneous alignment of three different data types. * Prioritized the maintenance of "World Knowledge" during multimodal training so that the addition of new modalities does not degrade the model’s factual accuracy. ## Enhancing Vocal Expressiveness and Prosody * Focused on "prosody"—the rhythm, pitch, and stress of speech—to move beyond robotic, flat text-to-speech (TTS) outputs. * Implemented a system of descriptive tokens and emotion tags (e.g., "warm voice," "excited tone") during training to give the model fine-grained control over its vocal persona. * Incorporated natural human speech elements, such as realistic breathing patterns and contextual variations in speech speed, to make interactions feel more intuitive and less synthetic. * Refined the model's ability to interpret the user's emotional state from their voice and respond with a matching emotional intensity. The evolution of Kanana-o highlights a shift from simply maximizing generic benchmarks to optimizing real-world user experiences through multimodal alignment and emotional intelligence. The success of this model underscores the necessity of high-quality, structured instruction data and fine-grained control over output styles to create truly conversational AI that feels natural to the user.