Sundar Pichai announced EmbeddingGemma 2, described as Google's first open, natively multimodal embedding model. It handles text, code, image, video, and audio tasks in a modular 740M parameter form factor, is positioned for offline, privacy-first RAG when paired with Gemma 4, and is said to outperform some specialist models more than twice its size. Weights are available now on Hugging Face.
French AI lab Mistral AI released Mistral Large 4 (ML4) on Tuesday, a 1-trillion-parameter multimodal model nicknamed Le Chonk. It is currently accessible only through a guarded public endpoint, with open weights planned within three weeks after safety testing is complete.
Mistral released its largest model to date, Mistral Large 4, with 1 trillion parameters and 49 billion active parameters. An API preview is available in Mistral Studio; weights are expected at the end of October.
Mistral released a public preview of Mistral Large 4, a natively multimodal model with 1 trillion parameters and 49 billion active parameters. Its weights will be opened at the end of the month.
Why it matters: Mistral outlined the open-weight release schedule for its 1-trillion-parameter multimodal model and its investment in European compute, allowing comparison with mainstream open models.
Mistral AI released Mistral Large 4, codenamed Le Chonk, with 1T total parameters, native multimodality, and 49B active parameters. The company says it is the best open-weight model on aggregated benchmarks in Europe and the U.S. It reaches state of the art on workloads in cyber defense, manufacturing, and finance, and exceeds closed frontier models in visual grounding. Built end-to-end in Europe, it can be deployed from Europe through Mistral Cloud. The API is available today, and open weights are due at the end of October.
Why it matters: Mistral specified a 1T-parameter, 49B-active open-weight model and a local European deployment path, offering a view of the open-weight field's scale and availability.
Reka AI released a research preview of Rho-1, a 19B-parameter general-purpose model that can process and generate text, images, video, and robot-control actions in a single neural network. Unlike most systems that route tasks to specialized models, Rho-1 places all modalities as tokens in one shared context window without calling tools or external models. It can generate continuous video in real time, respond instantly to new instructions without restarting, and use the same weights that predict camera images to drive robot actions.
Startups in NVIDIA’s Inception program are building AI applications to support breast cancer care, from imaging and risk assessment to treatment planning. iSono Health’s FDA-cleared ATUSA system captures a 3D breast ultrasound in approximately two minutes per breast; the company says it is 28% more sensitive than handheld 2D ultrasound. Ataraxis AI uses digital pathology slides and clinical variables to predict treatment response and recurrence risk.
AI 3D generation company Meshy said it grew annual recurring revenue from $1 million to $100 million in under two years, faster than HeyGen, which took 29 months to grow from $1 million to $100 million.
Researchers proposed an Image-to-Code method in which a coding agent iteratively writes executable Blender programs from a single photo, producing runnable, editable scene code. They also released the LEGO-Bench benchmark.
.@Cloudflare's decision models are available on Ollama! Decision models allow you to classify an image, label a bug report or route a support ticket to the right team. Clef (27B): ollama pull clef Clef Flash (9B): ollama pull clef-flash
Andrej Karpathy shared several ways to ask LLMs to make their outputs easier to understand: write using the ASD-STE100 controlled-language standard or relax it to 80%; generate diagrams; output an interactive webpage in HTML; or make a custom explainer video, for example with voiceover using an ElevenLabs API key. He thinks that as models improve, more work will move up to supervision and understanding, and models will be able to create large custom software artifacts that previously were not worth making.
Tavus launched Griffin, which the company calls the first Human Interaction Model (HIM). In live video calls, it can receive and generate speech, facial expressions, tone of voice, gestures, and pauses at the same time.
When AI-generated content becomes so realistic, how can we verify what was created by people, machines, or nature? In our podcast, we explore the critical role of provenance—and how imperceptible watermarks can protect both digital media and biological designs. ↓ Timestamps: 00:00 Introduction 00:34 What is watermarking? 04:35 SynthID 15:16 Images and video 19:28 SynthID Bio 28:00 A resilient future
Kaiming He's team proposed NAT-ARC, a purely visual ARC-solving system that does not rely on LLMs. It pretrains an encoder with MAE on ImageNet and transfers it to abstract grid reasoning. The best single model has 0.6B parameters and reaches 63.4% pass@2 on ARC-1; an ensemble reaches 70.2%, with about one-quarter the parameters of the specialized LLM system The ARChitects, which uses an 8B model and reaches 71.6%.
At its annual DevDay, OpenAI launched Dots, an agent powered by GPT-6 Astra that can be assigned professional tasks such as QA, website design, and trip planning. It is currently available only to subscribers on the $20-per-month Pro plan or above.
Google DeepMind released frontier model Gemini 4 Argon, initially making it available through the Fairwind Program to trusted cybersecurity defenders before gradually expanding access. Pricing starts at $2 per million input tokens and $10 per million output tokens.
Why it matters: Gemini 4 Argon raises the maximum output to 1M tokens and reports multiple benchmark results and internal deployment data, offering a view of its limits on long-running workflows.
GenFlow 1.0, launched jointly by Baidu Wenku and Baidu Netdisk, has been given the Chinese name Kuku AI and released as a standalone product. It enters tasks such as writing papers, making presentations, and generating industry reports as a general-purpose AI agent. The enterprise version has attracted 3,000 trial companies shortly after launch, with faster adoption among small and midsize businesses. Its international version has been refreshed as Kooko by Oreate AI and surpassed 10 million overseas users within a year.
QbitAI's interview program invited Gu Xin, who leads sales and ecosystem success for Siemens Xcelerator in China, and Huang Yao, founder and CEO of Aqiu Tech, to review the path from industrial AI deployment to scale. Huang used semiconductor wafer cutting as an example, saying that after introducing agent technology, equipment OEE and capacity can improve by about 25%. Gu said industrial AI must be reliable at an industrial level. Siemens Xcelerator aims to serve as a connector and amplifier, filling in the delivery and ecosystem pieces needed to scale from 1 to 100.
OpenAI released its next-generation flagship model GPT-6 Astra at 1 a.m. Beijing time on September 4. Leiphone compiled first-hand tests from users who received access early.
Why it matters: Early user tests show GPT-6 Astra can check its own work on long tasks but still needs human oversight, revealing its capability limits.
Aishi Technology fully launched PixVerse R2, available on the web for motion commands and interactive game-like experiences. R2 brings five growth dimensions—model capacity, data, tasks, control signals, and time scale—into one framework, advanced through two parts: the Omni Causal AR capability layer and the Real-Time Acceleration layer.
Using three photos of the bronze rat head from Yuanmingyuan and a natural-language instruction, GPT-6 Astra used Codex to direct Aholo Lux3D to generate a 3D asset, then used Blender to inspect it, adjust materials, add a display stand, and render and export the result.
Leiphone compared HiDream-O1-Video-1.0, Seedance 2.5, and MiniMax H3 on three original image-to-video tasks, evaluating intent planning, temporal causality, and audio-video alignment.
Tests found that Seedance 2.5 has stronger visual presentation and character consistency, but blocked a plot rewrite based on reference images from the series Journey to Love, first saying the images might depict public figures and then stopping generation over 'copyright-protected elements' after the references were removed. Kling VIDEO 3.0 generated a complete 15-second short film with a twist ending.
At an ECCV 2026 workshop, NVIDIA Research scientist Li Boyi presented two projects, MultiGen and DexImit. MultiGen uses diffusion models to add synchronized audio to videos generated by physical simulators.
In the first keynote at ECCV 2026, University of Texas at Austin professor Kristen Grauman argued that visual AI should move from understanding to enabling, becoming an AI Guide that helps people learn real-world skills.
ECCV 2026 received 10,473 submissions and accepted 2,834. The 3D topic fell from a leading area to third place, but questions around 3D representation, geometry, and physical consistency are spreading into video generation, embodied AI, and world models. Of 93 workshops, about 10 were directly related to 3D, 11 to spatial intelligence, nine to physical AI, and five to world models. Best-paper and honorable-mention awards also focused on 3D surface representation, object reconstruction, and geometry perception.
An AWS tutorial shows how to deploy two endpoints using the same vLLM-Omni DLC image on SageMaker AI: a real-time endpoint using FLUX.2-klein-4B for image generation and an asynchronous endpoint using Wan2.1-VACE-1.3B for image-conditioned video generation. A text prompt first produces a PNG, then the image and motion prompt are sent through Amazon S3 to the video endpoint. The generated MP4 is retrieved from S3. The example includes a command-line workflow and an optional Streamlit interface.
Former Yahoo CEO Marissa Mayer launched personal AI assistant Dazzle, which raised an $8 million seed round last December. Its context comes entirely from a phone’s camera roll rather than email or calendars. Dazzle can scan recent photos to schedule events or find a repair person, and can recommend vacation destinations and birthday gifts based on a photo library. Mayer said Dazzle prioritizes privacy and discards personal information that AI flags as sensitive.
AMD announced an all-stock acquisition of spatial intelligence startup World Labs for about $8.2 billion. The deal is expected to close by the end of 2026, subject to regulatory approval. World Labs founder Fei-Fei Li will join AMD's leadership team as executive vice president and chief scientist, reporting directly to CEO Lisa Su and leading frontier research.
Why it matters: Beyond the price and leadership changes, the source outlines the technical case for world models and compares the ecosystem with Nvidia's.
Condé Nast and AWS's Generative AI Innovation Center (GenAIIC) built a multimodal video search solution using Amazon Bedrock and Amazon OpenSearch Service. It uses TwelveLabs Marengo embeddings to jointly encode visuals, audio, and transcripts, cutting a single content search from 250 minutes to under two minutes.
Microsoft Research has released Quine, an AI research system designed for biological complexity. It combines a biological world model with an interactive harness connecting models, scientific tools, literature, and experimentalists. In work with the Broad Institute on pancreatic ductal adenocarcinoma, it predicted and prioritized thousands of compounds; the top-ranked compound produced the largest classical-to-ground-state transition in wet-lab experiments. The process from narrowing the search space to selecting a candidate for validation took one weekend.
Why it matters: Microsoft Research describes Quine's multimodal biological world model and experimental loop, including wet-lab validation in pancreatic cancer compound screening.
AMD acquired World Labs for $8.2 billion. The official blog says that since its founding in 2024, World Labs has built a model-training team for image, video, and spatial reconstruction, and has advanced robotics simulation through its acquisition of SceniX.
Why it matters: AMD's $8.2 billion acquisition of World Labs and its disclosure of Atlas's sparse-reconstruction progress offer a view into the integration of spatial intelligence and robotics simulation.
Latent Space's AINews roundup of developments from September 24–25 says community feedback was positive after Opus 5.5 launched that week, especially for generating explainer videos with code.
Why it matters: The weekly roundup offers a quick look at model and agent-infrastructure developments, including community feedback on Opus 5.5's code-generated videos and how vendors position their new models.
AMD announced an all-stock acquisition of Fei-Fei Li's World Labs for $8.2 billion. The deal is expected to close by the end of 2026, subject to regulatory approval. After closing, Li will join AMD as executive vice president and chief scientist, reporting directly to Lisa Su; co-founders Justin Johnson and Ben Mildenhall will continue leading the team as it joins AMD.
Why it matters: The full deal details and the companies' year-and-a-half collaboration are laid out, helping readers understand why a chipmaker would acquire a spatial-intelligence team.
In its first year, Microsoft Research Asia's Singapore lab (MSRA–Singapore), Microsoft's first research lab in Southeast Asia, has advanced work in four areas: next-generation AI models and agent systems, domain-specific AI, AI-native research practices, and ecosystem and talent development. The lab works with partners in Singapore's healthcare, finance, and education sectors to explore applications including multimodal medical AI and self-evolving diagnostic agents.
Using the open-source SkyRL framework on Amazon SageMaker HyperPod, GRPO post-training of the Qwen3-VL-8B vision-language model raised maze-solving success from 43.75% to over 95% on a fixed 64-maze evaluation set.
Runway has released a research preview of GWM Worlds 2, turning high-fidelity video and audio generation into a real-time interactive simulation. It introduces WorldPrompt, an input format that can hold parts of the environment—including the first frame—constant while generating timestamped events that can also be prompted in real time.
Google Research has introduced an AI video co-director multi-agent framework, an orchestration layer on top of Gemini and Veo. It models long-form video generation as a problem of global optimization and world-state tracking to mitigate semantic drift and cascading failures.
Why it matters: Google breaks long-video generation into four composable frameworks and reports continuity improvements across several benchmarks, illustrating how multi-agent orchestration can reduce semantic drift.
Google DeepMind has released Gemini 3.8 Live with Live Avatar, combining real-time video generation and voice to give Gemini's conversational model a near-real-time visual avatar, with accurate lip-sync, natural expressions, and smooth turn-taking.
Why it matters: Gemini 3.8 Live adds real-time video avatars, giving readers a view of how multimodal conversation is being applied in enterprise settings and where the feature is available.
Liquid AI has released the experimental DSpark draft model for its LFM2.5-VL-3B vision-language model. Speculative decoding speeds inference without changing output quality: up to 3.13× on-device and 2.66× on an H100, with end-to-end speedups of up to 2.62× and 2.27×, respectively.
