Simon Willison's LLM CLI tool and Python library released version 0.37, introducing a command to easily configure OpenAI-compatible endpoints. The update switches the default model to the cheaper gpt-6-luna and enhances embedding support for multimodal binary content including text, audio, and images. The author also notes that while the project receives many AI-assisted pull requests, large language models are effectively used to triage and review these contributions.
Multimodal Embedding
多模态嵌入
Mapping different data modalities into a shared vector space representation.
Curated Intelligence
Amazon demonstrates a multi-agent system combining Nova Multimodal Embeddings and Bedrock AgentCore to automate part substitution decisions in aerospace. By retrieving data from technical diagrams and specifications, the system reduces days of manual cross-referencing into auto-generated assessment drafts, significantly improving engineering review efficiency.
Google released EmbeddingGemma 2, an open-weight multimodal embedding model based on Gemma 4 that maps text, code, images, video, and audio into a unified 768-dimensional vector space. Optimized for on-device deployment under the Apache 2.0 license, it features Matryoshka Representation Learning (MRL) to balance storage efficiency and retrieval quality.
Developers repurposed Google's multimodal embedding model EmbeddingGemma 2 into a high-speed local speech-to-text tool, transcribing one hour of audio in approximately 13 seconds on a base MacBook Air. The approach freezes the original audio encoder and trains only a lightweight 'reader' to map audio slices into the text space, running entirely on the Apple Neural Engine with no data leaving the device.
Google has released EmbeddingGemma 2, an open-source multimodal embedding model based on the Gemma 4 architecture that unifies processing of text, code, images, video, and audio. By leveraging quantization to reduce memory footprint to just 191MB, it significantly lowers the barrier for deploying high-performance vector retrieval and semantic understanding on mobile and edge devices.
Google DeepMind has released EmbeddingGemma 2, an open-source lightweight model built on the Gemma 4 architecture that maps text, images, audio, and video into a unified embedding space. Designed for consumer hardware, it enhances on-device search and multimodal RAG pipelines, building on the predecessor's success with over 20 million downloads.
Google DeepMind launched EmbeddingGemma 2, upgrading its previous text-only model to support unified embeddings for code, images, video, and audio. Built on the Gemma 4 architecture, it targets lightweight, privacy-first RAG pipelines on consumer hardware. The model is now open-sourced with guides for inference and fine-tuning.
LangChain has released a multi-modal RAG template specifically designed for slide decks, addressing the limitation of traditional text-based RAG in handling visual information. By integrating image understanding with vector retrieval, this solution enables LLMs to extract context and generate answers directly from unstructured visual data like presentations, with benchmarking and deployment support available via LangSmith.
The Sentence Transformers library now supports training and fine-tuning multimodal embedding and reranker models across text, image, audio, and video. Using Qwen3-VL-Embedding as an example for Visual Document Retrieval, this update lowers the barrier to building customized multimodal RAG systems.
The Sentence Transformers library released v5.4, officially introducing multimodal encoding and comparison capabilities for text, images, audio, and video. Developers can now build cross-modal Retrieval-Augmented Generation (RAG) and semantic search applications via a unified API, significantly simplifying the deployment of multimodal embedding and reranker models.