AI Briefing
KOSign in

Kakao Releases Kanana-2-vision-encoder, Open-Sourcing Medium and Tiny Weights

·2026.10.21 08:00

Key point

Kanana-2-vision-encoder recorded the global highest scores in 8 out of 9 domains.

1 / 16

Details

  • Initialized with Kanana-2 LLM weights to boost training efficiency by 4x, and reduced latency by 61.8% when processing high-resolution images using SWA and GQA structures
  • Recorded the global highest scores in 8 out of 9 benchmark domains, with a significant lead in visual spatial awareness and Korean OCR
  • Built a parameter lineup of Large (900 million), Medium (400 million), Small (200 million), and Tiny (100 million), and open-sourced the Medium and Tiny weights under the Kanana-open license
  • Applied pruning, Full-Vocabulary KL distillation, and On-Policy Distillation to ensure precise visual understanding performance even in smaller models
open-sourcemultimodalvision-encoderknowledge-distillationkakaokanana-2-vision-encoder
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditAgnes AI Releases Agnes-3.0-Qwen: 180B MoE Model with 1M Context and Strong Agentic BenchmarksRedditinclusionAI Releases Ming-3.0-audio-tiny, a 9.3B Parameter Native Audio-Language ModelKakaoKanana-2 Development Log (1): Focusing on Decision-Making in the Pre-training StageKakaoKakao Releases Open-Source 'Kanana-2' Optimized for Agentic AIGeekNewsArtCraft releases open-source IDE for AI image and video creation with 62-model catalogDeepSeekDeepSeek Releases V4.1-Flash-1017 and Opens Search & Fetch APIs