Google engineers demonstrate how AI systems are evolving from basic retrieval-augmented generation (RAG) to graph-based architectures that handle more complex reasoning and multimodal tasks. The 90-minute workshop walks through building production agent stacks with semantic graph retrieval and specialized agent orchestration.

• RAG is being replaced by graph-based approaches that organize context semantically rather than just retrieving relevant documents
• The production stack involves extracting graph context and orchestrating multiple specialized agents to handle different tasks
• This represents a concrete shift in how companies are building AI systems—from simple retrieval to structured knowledge representation

MiniMax H3 is a general-purpose AI that takes text, images, video, and audio as input to generate up to 15-second, 2K videos with native stereo sound. It matches or beats mainstream models on price-performance, excels at instruction following and brand rendering, and will release its weights soon under open-source terms.

• MiniMax H3 unifies text, image, video, and audio into one model that generates 15-second 2K videos with native stereo sound, costing under one-third the price per second at 2K versus mainstream models.
• It can follow complex multimodal instructions, merging camera movement, character appearance, and audio from separate source inputs into one coherent output.
• Four new technologies—Contextual Omni Representation, H3-VAE, Omni Transformer, and In-Context Regeneration—drive its compression, throughput (+30%), and cross-modal coherence.
• MiniMax plans to open-source the model weights, designed for broad hardware compatibility, aiming to seed a wider ecosystem.

The article introduces Inkling-Small, a Mixture-of-Experts transformer with 276 B total and 12 B active parameters that matches full Inkling performance at one-quarter the size. It handles text, images, and audio with a 1 M-token context window and lets users trade compute for thinking effort. The open-weight model outperforms peers on tool use, coding, reasoning, and multimodal benchmarks and is available for fine-tuning and chat.

• Inkling-Small matches or beats full Inkling performance (e.g., 31.6% vs 29.7% on Humanity's Last Exam) while using only 12B active params out of 276B total—roughly a quarter the compute cost.
• It's natively multimodal (text, images via 40×40 patches, audio via dMel spectrograms) with a 1M-token context window and adjustable "thinking effort" for cost/performance tradeoffs.
• Safety numbers slightly lag full Inkling (71.6% vs 78.0% harmful-query refusal on FORTRESS) despite reusing the same safety recipe.
• Open-weight and available now for fine-tuning on Tinker and chat testing in the Tinker Playground.

This article introduces the Gemma 4 family of models from Google DeepMind, detailing their architectures and improvements over the previous version, Gemma 3. It highlights key features such as interleaved attention layers and efficiency enhancements in global attention mechanisms.

• Gemma 4 comes in four sizes—E2B, E4B, 31B, and a 26B MoE model that only activates 4B parameters at inference—and handles text, images, and audio
• Local-to-global attention layers are interleaved at a fixed 4:1 ratio in smaller models and 5:1 in larger ones, with sliding windows of 512 tokens (E2B/E4B) or 1024 tokens (larger models) to cut compute costs
• Global attention memory use is reduced through Grouped Query Attention, a K=V trick that makes keys and values identical, and p-RoPE for improved positional encoding

This article explores how advanced AI models can generate detailed image descriptions and reasoning without actual image input, a phenomenon called mirage reasoning. It highlights vulnerabilities in these models, particularly in medical contexts, and introduces B-Clean, a method for better evaluating multimodal AI systems by minimizing non-visual inference.

• A model topped a chest X-ray QA benchmark while being completely blind to the actual images, exposing "mirage reasoning" where models fabricate plausible visual descriptions from text priors alone.
• Explicitly prompting models to guess without assuming an image was present caused performance to drop sharply, showing benchmark scores were inflated by non-visual inference rather than real image understanding.
• The proposed B-Clean method strips out textual cues that let models cheat their way to correct answers, aiming to force genuine visual grounding in evaluations—critical for high-stakes fields like medical AI.

Qwen has released the Qwen3-VL-Embedding and Qwen3-VL-Reranker models, designed for advanced multimodal information retrieval and cross-modal understanding. These models support various inputs, including text and images, and enhance retrieval accuracy through a two-stage process of initial recall and precise re-ranking.

• Qwen3-VL-Embedding and Qwen3-VL-Reranker pair up for a two-stage retrieval pipeline (recall then rerank) covering text, images, screenshots, and video across 30+ languages.
• Embedding model uses a dual-tower design for independent encoding, while the reranker uses a single-tower architecture with cross-attention for deep query-document interaction.
• Achieves state-of-the-art results on image, visual document, and video retrieval benchmarks.
• Trails the text-only Qwen3-Embedding model on pure text retrieval, showing a tradeoff for its multimodal gains.

Multimodal vector databases like ApertureDB are revolutionizing how industries manage and verify data, particularly in healthcare advertising. By integrating various data types and employing AI tools, these databases enhance compliance by detecting omissions in marketing content, ensuring that critical information is accurately conveyed to patients.

• ApertureDB combines multimodal vector search with AI to flag missing required information (like side effects or risks) in healthcare marketing content.
• The system helps compliance teams catch omissions before ads reach patients, reducing regulatory and safety risks.