LiquidAI has introduced the LFM3 framework to enable developers to build specialized multimodal large models optimized for edge devices. The tool focuses on efficient inference and custom deployment in resource-constrained environments, currently in an early community sharing phase.
Multimodal Model
多模态模型
An AI model capable of processing multiple data types like text and images simultaneously.
Curated Intelligence
An InfoQ and QCon panel featuring tech leaders from major firms discussed the scalability bottlenecks of AIGC video generation. While single 10-second clips are impressive, delivering full short dramas faces challenges in production stability, human-AI collaboration boundaries, and true cost control. Experts noted that AI currently assists with plot expansion and structure, but thematic expression and emotional control require human oversight, with multi-scale information alignment being a key technical hurdle for long contexts.
Liquid AI has open-sourced d1-3B and the experimental multimodal d1-omni-600M from its d1 decision model family, optimized for edge computing. The d1-3B model achieved a score of 48.57 on the Decision Index 0.2.1, establishing it as the best-performing decision model under 10 billion parameters, outperforming several 4B, 9B, and even 35B-scale competitors.
Google launched Playground, a browser-based AI platform enabling US users to create and test games via text prompts without coding skills. Integrating Gemini, Nano Banana, and Lyria models, it allows real-time tweaking of rules and physics, complementing professional tools like Unity Spark.
Meshy ranked 31st in a16z's inaugural consumer AI app monthly revenue list, standing as the only AI 3D company on the chart. Based on verified bank card spending data from YipitData, this milestone confirms AI 3D generation has emerged as a distinct and monetizable vertical category.
French AI lab Mistral released Mistral Large 4 (codenamed Le Chonk), a 1-trillion-parameter multimodal model aiming to challenge US and Chinese rivals. Currently accessible only via a guarded endpoint, the weights are scheduled for open-source release in three weeks after safety testing, positioning the model as a 'third way' between closed and open ecosystems.
Meta has released and open-sourced Muse Glimmer, a 30B-parameter multimodal model distilled for local agentic use cases under the Apache 2.0 license. Designed to balance privacy and cost efficiency, it supports coding, document analysis, and personal assistants, with immediate Day-0 support across the Hugging Face ecosystem including Transformers and vLLM.
NVIDIA introduces Nemotron 3 Nano Omni, a long-context multimodal model designed for document, audio, and video agents. It aims to enhance AI agent understanding and reasoning in complex scenarios by efficiently processing large-scale unstructured data.
OpenAI introduced a faster ChatGPT Images experience in 2025 and launched the GPT-Image-1.5 model via its API. Users can now explore the latest ChatGPT Images 2.5.
The Hugging Face team published an article systematically reviewing the technical logic and ethical challenges behind the rise of AI agents. It notes that as LLM capabilities advance rapidly, the industry is shifting toward autonomous agents, exploring their training paradigm as 'multimodal model++' and potential social risks. The review aims to provide developers with an ethical framework and reflective perspective for building agents.