AI Briefing#multimodal
The latest AI and developer news about #multimodal, with the original source and a short summary.
The latest AI and developer news about #multimodal, with the original source and a short summary.
A tool has been released that supports multimodal fine-tuning of Gemma models, including text, image, and audio, on Apple Silicon environments.
Alibaba has unveiled Qwen3.6-Plus, which significantly enhances coding and multimodal agent performance.
Google DeepMind has released the new Gemma 4 series of open models supporting vision and audio.
Gemma 4 has been released with multimodal, audio, and on-device support, along with 128K/256K context.
A 0.6B early-fusion model unifies open-vocabulary grounding and segmentation.
Z.ai's GLM 5V Turbo is now offered on Vercel AI Gateway.
Chinese AI startup **DeepSeek** experienced an unusual service outage that lasted for hours.
As Meta delays the launch of its next-generation AI **Avocado**, it is narrowing the performance gap through testing of various variant models and parallel use of **Gemini**.
Qwen3.5-Omni significantly boosts omnimodal performance with a 256k context, 10-hour audio input, and 113-language ASR.
Based on Nova 2.0 Lite, attaching LoRA to only **o_proj** gave the best balance of efficiency and accuracy.
OpenAI has released GPT-5.4 mini and nano, high-performance small models optimized for coding and subagent tasks.
Forge trains, aligns, and evaluates frontier-level AI models using a company's internal knowledge.
Mistral Small 4 has been released, combining reasoning, multimodality, and coding into one model.
Kanana-V is expanding to include understanding of Korean culture and screen control capabilities.
Qwen 3.5 Plus, optimized for agentic workflows with a 1M context window, has launched on AI Gateway.
Qwen3.5-397B-A17B has been released, a native multimodal model in which only 17B of the 397B parameters are activated.
China's open-source AI ecosystem is evolving through MoE architecture, multimodal expansion, and small model optimization.
The first development log for Kanana-2 summarizes the key decisions made during the pre-training stage.
The Kakao Kanana organization has released 'Kanana-v-embedding', a 2-billion parameter multimodal embedding model, demonstrating superior search performance compared to Jina Embeddings v4 through training on the Korean-specific dataset KoEmbed.
**Qwen3-Omni-Flash-2025-12-01** has been upgraded to handle speech, vision, and text more reliably across the board.
AutoGluon Assistant uses multi-agent LLMs to automate model training without writing any code.
HuggingFace's TRL library now supports the latest alignment techniques such as MPO and GRPO for VLMs.
This proposes a method for building a data pipeline that minimizes padding waste to maximize multimodal model training efficiency.
Google's multimodal model Gemma 3n has been released through various open-source libraries.