Kakao updates Kanana-2, expands open-source push with 4 new AI fashions

Effectivity-first design targets agentic AI adoption past huge tech corporations

Kakao mentioned on Jan. 20 that it has upgraded its proprietary massive language mannequin, Kanana-2, and launched 4 further variants as open supply, extending its effort to place the mannequin as a sensible basis for agentic synthetic intelligence. The replace follows Kakao’s preliminary choice to open-source Kanana-2 in December through Hugging Face, the place the mannequin attracted consideration for combining comparatively sturdy efficiency with decrease computational calls for.

Kakao framed the replace as a transfer to enhance usability somewhat than merely scale dimension. The corporate mentioned the revised fashions are designed to run effectively on extensively used {hardware} similar to Nvidia A100-class GPUs, making them accessible to small and midsize companies, startups, and tutorial researchers that will lack the sources required for cutting-edge infrastructure.

Architectural modifications aimed toward effectivity

On the core of the replace is a stronger deal with compute effectivity. Kanana-2 makes use of a mixture-of-experts (MoE) structure, through which solely a subset of the mannequin’s parameters is activated throughout inference. Whereas the mannequin has a complete of 32 billion parameters, Kakao mentioned solely round 3 billion are used at a time, considerably reducing computational price and not using a corresponding drop in efficiency.

Kakao additionally launched a brand new “mid-training” section between pretraining and post-training. This extra stage is designed to present the mannequin new reasoning talents whereas preserving present data, addressing the frequent downside of fashions dropping earlier capabilities as they study new duties.

4 fashions for various use instances

The up to date open-source launch contains 4 variants, every focusing on a particular kind of use:

• Base: a general-purpose basis mannequin
• Instruct: tuned for instruction-following duties
• Considering: optimized for reasoning-heavy workloads
• Mid-training: aimed toward analysis and continued adaptation

Kakao mentioned all 4 fashions emphasize price effectivity whereas strengthening tool-calling capabilities, a key requirement for agentic AI methods that must act, not simply reply.

Deal with agentic AI, not chatbots

In contrast to standard conversational fashions, Kanana-2 is designed to assist agentic AI, which might interpret advanced person requests, resolve which instruments to make use of, and execute multi-step actions. Kakao mentioned the fashions have been skilled on intensive multi-turn datasets centered on device utilization, enabling them to autonomously choose and function exterior instruments in real-world situations.

This design alternative displays a broader shift within the AI subject, the place builders are more and more centered on methods that may carry out duties—similar to retrieving information, executing workflows, or interacting with software program—somewhat than solely producing textual content.

Benchmark outcomes and positioning

In benchmark testing, Kakao mentioned the up to date Kanana-2 fashions outperformed a peer open-source mannequin, Qwen-30B-A3B-Instruct-2507, in instruction-following accuracy, multi-turn tool-calling, and Korean-language duties. Whereas Kakao didn’t place the outcomes as a definitive rating, it argued that the comparisons present efficiency-focused fashions can stay aggressive on core capabilities.

Trade observers notice that such positioning displays a rising emphasis on “adequate” efficiency mixed with decrease price, somewhat than pursuing ever-larger fashions at greater expense.

Reducing obstacles to AI adoption

One other central theme of the replace is accessibility. Kakao mentioned Kanana-2 is optimized to function on general-purpose GPUs somewhat than requiring the newest high-end accelerators. This strategy is meant to scale back obstacles for organizations that need to deploy superior AI however face price range or infrastructure constraints.

Kim Byung-hak, efficiency lead for Kakao’s Kanana challenge, mentioned the objective was to construct sensible agentic AI with out counting on pricey {hardware}. “By open-sourcing high-efficiency fashions that work on common infrastructure, we hope to supply a brand new various for AI adoption and assist the expansion of Korea’s AI analysis ecosystem,” he mentioned.

Kakao added that it’s persevering with growth of a bigger 155-billion-parameter mannequin, signaling that it’s pursuing each efficiency-oriented releases and longer-term ambitions to compete on the high tier of worldwide AI growth.