#多模态

Introducing Forward Deployed Creatives in ElevenCreative

ElevenLabs introduces Forward Deployed Creatives, creative professionals who work with teams in ElevenCreative or produce finished creative end to end, addressing brand consistency, fragmented tools, and creative control. Enterprises can enable their own creative team on the platform, build reusable workflows with Forward Deployed Engineers and Creatives, or extend their team with end-to-end creative services from strategy to localized ads. Localized content is reviewed by trained linguists and audio professionals, and ElevenLabs also offers end-to-end production services for dubbing, voiceovers, subtitles, audio description, and custom brand voices.

Scaling Intelligence: A landmark commission and exhibition

Modal and Gray Area opened Scaling Intelligence, an exhibition of 11 new artworks made with machine intelligence and cloud-scale compute, with Modal supporting the artists through engineering help and compute credits. The artists were selected through a global open call for practices that use AI with intention and imagination, and their works include a satellite searching for a nonexistent island, a robot arm that flips off passersby, and a choir of nine AI models that learn from visitors singing along.

Microsoft Research Focus – Edition #98

This edition of Microsoft Research Focus covers AI-powered universal translation, the resolution of the 31-year-old KLS conjecture in high-dimensional geometry, and a new approach to making AI coding agents faster and more efficient. It also highlights a research intern's journey at Microsoft Research Asia – Tokyo and the Latent Core Tokenizer, which produces more linguistically meaningful tokens across 104 languages.

Inkling 2 Small on vLLM: Day-0 Support

vLLM announces day-0 support for Thinking Machines' Inkling 2 Small, a multimodal text-and-vision MoE model with 594B total and 17B active parameters, up to 1M tokens of context, and a native DSpark speculator. The post details the architecture (48 layers, 384 experts with top-8 routing, block attention residuals, gated attention, a 23-layer ViT image encoder, and mixed NVFP4/MXFP8 expert precision) and vLLM's optimizations, including kernel fusion with PDL, fused MoE activation quantization, sequence parallelism for prefill, and a Gluon relative attention kernel, reaching up to 630 TPS/user at TP4 on GB200. It also covers LoRA and RL (via SkyRL) support, a serve command, agentic-workload benchmarks, and accuracy checks on NIAH up to 1M tokens, MMMU-Pro and HLE.

Gemini Decisions API: Fast, calibrated multimodal decisions

Google introduced the Gemini Decisions API, powered by the new Gemini Decision 1 model (gemini-decision-1) and available through the Gemini API in Google AI Studio, for closed-set decisions such as tool selection, policy checks, model routing, and context grading. It accepts one multimodal input (text, images, audio, video, PDFs up to 256K tokens) with up to 1,000 yes/no, multiple-choice, and rubric-scoring questions, scoring options in a single forward pass with confidence scores at 150-250 ms end-to-end latency and $0.10 per 1M input tokens with zero output tokens billed. Google reports that Gemini Decision 1 leads across agent routing, reasoning, ranking, and multimodal benchmarks with the lowest calibration error, and is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API across the shared benchmark suite.

Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD

vLLM details its Encoder/Prefill/Decode (EPD) disaggregation work, including metadata-only requests that let downstream instances reuse remote embeddings without repeating media preprocessing, a dedicated encoder-only Model Runner V2 path, and remote embedding transfer via NIXL and Mooncake. It also covers cross-encoder output reuse through Mooncake Store and orchestration support in vLLM Router, SMG, and llm-d, with core capabilities in vLLM v0.32.0 and DeepSeek V4.1 Flash metadata-only support on main via #62802. A DeepSeek V4.1 Flash benchmark on GB200 found EPD delivered 6.6–9.5% higher per-GPU throughput and 44–47% lower P90 TTFT than PD at concurrency 64–128, but lower per-GPU throughput at concurrency 16.

Experience the art and legacy of Dr. Esther Mahlangu

Google Arts & Culture has launched the first online retrospective of South African artist Dr. Esther Mahlangu, timed to her 91st birthday, featuring more than 45 curated stories and an archive of her art, life and legacy. Curator Nontobeko Ntombela describes using AI, including Veo, to visually recreate Mahlangu's undocumented memories from age 10 until her first documented works in 1980, in collaboration with Google DeepMind, the Google Arts & Culture Lab and Domestic Data Streamers. The exhibition was curated with partners including Art of Contemporary Africa, Iziko Museums of South Africa, the British Museum, the Jean Pigozzi Collection, the National Museum of Women in the Arts and Serpentine, and is available at goo.gle/esther-mahlangu.

Seedream 5.5 is now available via Partner Nodes

Seedream 5.5 is now available in ComfyUI via Partner Nodes as a multimodal model that generates from text and edits from images. The release adds multi-turn localized editing, native transparent RGBA output, poster and infographic layouts with readable text, more realistic portraits, and prompt-based lighting control, all behind a single node. Partner Nodes run on credits, and users can load the Seedream 5.5 template from the Templates panel or use Comfy Cloud.

Multimodal Knowledge Agent (MKA) benchmark: testing knowledge work capabilities with complex visual files

Artificial Analysis partnered with Google DeepMind to launch the Multimodal Knowledge Agent (MKA) benchmark, comprising 200 private real-world visual tasks across 10 industries in which agents work from dense visual files like architectural drawings, engineering schematics, and clinical scans to produce deliverables such as slide decks, spreadsheets, PDFs, Word documents, and marked-up images. Tasks are graded against expert rubrics averaging ~70 weighted criteria split into Regression and Climb checks, with the headline Climb WSR score, run in the open-source Stirrup agent harness with up to 250 turns and graded by GPT-6 Luna (medium) as an agentic judge. Claude Fable 5.2 (max) leads with 57% Climb WSR, followed by Gemini 4 Carbon (high) at 54% and Claude Opus 5.5 (max) at 53%, while Claude Fable 5.2 averages $31.97 per task versus $6.29 for Gemini 4 Carbon and $1.11 for GPT-6.1 Sol.

Introducing Scribe v3 and v3 Live, our most accurate speech to text models

ElevenLabs launched Scribe v3 for recorded audio, which makes 27% fewer word errors than Scribe v2 across 102 languages and has the lowest diarization error of any model tested, and Scribe v3 Live for realtime audio, which makes 43% fewer semantic word errors than Scribe v2 Realtime and adds speaker diarization and improved turn detection. Scribe v3 Live is available today in ElevenAgents, and both models are available in ElevenAPI, where switching from Scribe v2 is a one-line change to the model ID (scribe_v3 or scribe_v3_live). Both run in the US by default with isolated EU, India, and Singapore environments, Zero Retention Mode, and SOC 2, ISO 27001, PCI DSS Level 1, and HIPAA-eligible coverage.

Best OCR Model for Scanned Documents, Measured

The author tested 15 vision models plus Mistral OCR and Tesseract on eight synthetic scanned pages scored by character error rate and 264 key-value checks, finding Claude Opus 5.5 most accurate (262 of 264 at $16.42 per 1,000 pages) and Mistral OCR the best value (258 of 264 at $2.02). The post documents failure modes such as fabricated card digits, wifi passwords and server names on a faded receipt, and provides code for transcription, disabling reasoning, cross-checking two models from different labs, and using the mistral-ocr engine.

We’re bringing Adobe’s creative and productivity tools closer to your everyday Google apps.

At Adobe MAX, Google announced an expansion of its partnership with Adobe that will bring Adobe's creative and productivity tools into Google products, and Google's frontier AI models into Adobe's apps. Planned integrations include a direct pathway from Google Drive and Gmail into Adobe Acrobat with an "Edit with" entry point in file previews, expanded Adobe connectors in Gemini and Gemini Enterprise, Adobe creative tools in AI Mode in Google Search, and Adobe's CX Enterprise Coworker agent in Gemini Enterprise. Adobe and Google DeepMind will also explore new editing capabilities for creative professionals, starting with an image model for Photoshop that could enable near-real-time edits.

Exploring the world's languages with Language Explorer

Google introduced Language Explorer, an interactive web platform developed as part of its 1,000 Languages Initiative, which lets researchers, educators, and language enthusiasts discover and visualize linguistic diversity around the world using data across thousands of languages built on the open-source LinguaMeta dataset. The tool is featured as an interactive installation in the new exhibition "Real Talk! Language. Power. World." at the Museum of Communication Berlin, running through February 2028, with a companion exhibition at the Museum of Communication Frankfurt through September 2029. Google also noted that Google Translate now supports nearly 250 languages and Gboard supports 900+ language varieties.

NVIDIA Nemotron 3.5 Super VL is now available in Amazon SageMaker JumpStart

NVIDIA Nemotron 3.5 Super VL, a 120B-parameter vision-language model with 12B active parameters, a C-RADIO vision encoder, and up to 1M-token context, is now deployable in Amazon SageMaker JumpStart via SageMaker Studio or the SageMaker Python SDK. JumpStart offers an NVFP4 variant on G7 instances and BF16 reference weights on ml.p5en.48xlarge, with measured throughput of 63.07 tokens/sec on ml.g7.24xlarge and 140.85 tokens/sec for BF16 on ml.p5en.48xlarge. The model targets document intelligence, GUI understanding, video question answering, and multimodal agentic applications.

A single AI model simulates the behavior of many substances, from sliding sand to folding proteins

Researchers at MIT, UBC and elsewhere have developed WorldParticle, a deep-learning system that represents materials as Lagrangian particles and predicts their motion one step at a time, learning underlying dynamics rather than relying on hand-coded solvers. The model was demonstrated on six physical phenomena, including deformable objects, granular materials and proteins, with higher accuracy than numerical methods and modeled protein structural changes about 100 times faster than numerical simulators. The work will be presented at ACM SIGGRAPH Asia.

Ads Engine adds human review from localization experts and creative specialists

ElevenLabs Ads Engine now lets teams have ElevenLabs localization experts and creative specialists review ads made in the tool, including new variants and localizations for markets where no team member speaks the language. Ads Engine connects to ad platforms, pulls existing creative, translates text, adapts images and dubs video, with linguists from ElevenProductions and in-house creative specialists reviewing before ads are finalized. It uses Brand Kit to apply colors, fonts, logos, reference images and rules across variants, and can return finished creative for approval or push it directly to ad platforms.

VIALS-AA: different paths to scientific image reasoning

VIALS-AA analyzes how 19 models reason over scientific images, classifying 9,177 tool calls from its original test set: 38.3% were image viewing and 47.3% prepared or measured images. Image measurement made up just 5–6% of tool calls for GPT-6 Astra (max) and GPT-6.1 Sol (max), versus 25–27% for Claude Opus 5.5 (max) and Claude Sonnet 5.5 (max). The report also notes that Gemini 3.8 Flash (high) accounts for 86.7% of commentary/no-op calls, that more intelligent models tend to call fewer tools, and that cost per task varies by nearly 100× across models.

VeRL-Omni v0.3.0: Async Training and Broader Multimodal Post-Training

VeRL-Omni v0.3.0 adds asynchronous and decoupled RL execution paths across multimodal workloads, makes the TransferQueue/ReplayBuffer-based V1 trainer the default, and moves the rollout stack to vLLM 0.30.0 and vLLM-Omni 0.30.0. It includes end-to-end recipes for MiniMax H3 audio-video post-training with DiffusionNFT and MiniCPM-o 4.5 AVQA GSPO, where a 4×H800 experiment found separate-async about 9% slower per prompt at natural response lengths but roughly 6.7% faster at 1024 tokens. The release also expands model and algorithm coverage, adds Ascend and AMD ROCm Dockerfiles, and notes upgrade changes including V1 as default, named diffusion media artifacts, and fail-closed attention backends.

Clearing up a robot’s fuzzy memory

MIT researchers developed UQ-DAAAM, a memory system that quantifies a robot's uncertainty about object descriptions by comparing multiple viewpoints and fusing them into more accurate captions, resolving ambiguities for about 92 percent of objects with only two additional captions each. The system improved object description accuracy by 5 percent over the state of the art and produced more accurate answers about objects, their locations, and when they were seen. The work, led by senior author Luca Carlone with co-lead authors Harry Zhang and Nicolas Gorlo, will be presented at the Conference on Neural Information Processing Systems.

New NVIDIA Cosmos 3 Open Models for Interactive World Simulation

NVIDIA's Cosmos world model platform adds four new open-weight checkpoints to its Cosmos 3 family: Cosmos3-Nano-Transfer-Auto for multi-view driving data with varied weather, lighting and time of day; Cosmos3-Nano-Sim-Auto for streaming synchronized camera and LiDAR driving simulation; Cosmos3-Nano-Sim-Transfer for turning simulator controls and text prompts into photorealistic video; and Cosmos3-Nano-Sim-Bimanual for predicting video from robot or camera actions. All four are post-trained examples of Cosmos 3 Nano with 16B trainable parameters and use OpenMDW1.1, permitting commercial and non-commercial use under its terms. The post also details a four-stage training recipe (bidirectional adaptation, teacher forcing, consistency distillation, self-forcing) and post-training guidance for adapting the models to new sensor rigs and embodiments.

Frameboy: Atlas in Real-time

The team behind Atlas, an autoregressive world model that generates frames one at a time with pixel-perfect camera control, released Frameboy, a research preview that streams frames from cloud GPUs to a browser at up to 15 fps as the user controls the camera. Frameboy serves a highly optimized Atlas variant, and the real-time capability is also integrated into Viewfinder, a web app (currently in closed beta) for building scenes frame-by-frame and exploring them live via a streaming Frameboy overlay. The post argues that faster generation lowers the cost of a "bad idea" and shifts interaction from cautious manual prompting toward intuitive exploration.

How Amazon Marketplace evaluates content quality with Amazon SageMaker AI

Amazon describes a system that renders a full A+ Content page as a single snapshot and scores it with a fixed Vision-Language Model across four rubric dimensions, then fuses those scores with four OCR readability signals in a CatBoost classifier trained on expert annotations. On a 20 percent holdout set of 5,478 product detail pages from 10 categories, the fused model reached 0.94 accuracy and 0.95 F1, versus 0.80 and 0.84 for an averaged-score baseline, and SHAP values plus an LLM turn the results into prioritized recommendations. The pipeline is implemented on Amazon SageMaker AI using Processing Jobs, Inference endpoints, Feature Store, Training Jobs, Model Registry and monitoring, with a diffusion model generating improved image assets.

Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips

Google DeepMind released Gemini Omni 1.2 Flash, now generally available across the Gemini API and Gemini for business, adding native 15-second video generation, multi-turn conversational editing, and stronger reference adherence. The model is priced 30% below Gemini Omni 1.1 at $0.07 per second for 720p video, with duration settable in one-second increments from 3 to 15 seconds. It also adds more natural speech and lip-sync, more controllable cinematic camera moves, and outputs carrying an imperceptible SynthID watermark.

Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips

Google DeepMind released Gemini Omni 1.2 Flash, now generally available across the Gemini API and Gemini for business, adding native 15-second video generation, multi-turn conversational editing, and stronger reference adherence. The model is priced 30% below Gemini Omni 1.1 at $0.07 per second for 720p video, with duration settable in one-second increments from 3 to 15 seconds. It also adds more natural speech and lip-sync, more controllable cinematic camera moves, and outputs carrying an imperceptible SynthID watermark.

Featured coverage of this story: 《Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips》

Detecting signs of hypertension using wearable sensors

Google researchers describe AI-PPG-ACC-HTN, a multimodal AI system that uses smartwatch PPG and accelerometry data, without cuff calibration, to detect a persistent hypertensive state by aggregating 15-second waveform embeddings over multiple days alongside height. Validated against 24-hour ambulatory blood pressure monitoring in a 196-participant, five-site US study, it reached 65.8% sensitivity and 90.0% specificity for initial screening and 78.4% sensitivity, 90.9% specificity and 96.7% PPV for confirmatory testing among 48 participants with elevated office BP. The authors frame it as early feasibility research for passive opportunistic screening, not a replacement for clinical cuffs or a standalone diagnosis.

Redesigned Health app expands language support

Apple's redesigned Health app expands language support to global English, Chinese (Simplified and Traditional), French, Japanese, Korean, Norwegian, Portuguese, Spanish, Turkish, and Vietnamese. Built with Apple Intelligence, it adds a Longevity tab with a Health Age feature derived from Apple Watch metrics and at-home movement evaluations that run entirely on device. The app will be available on Apple Intelligence-enabled iPhone and iPad models with the latest software, with some features limited by region, language, or device.

Introducing India-based inference for Kimi K3 on Amazon Bedrock

Amazon Bedrock now supports in-country inference in India for Moonshot AI's multi-modal, open-weight Kimi K3 model via the in.moonshotai.kimi-k3 geographic inference profile, processing requests in the Mumbai and Hyderabad Regions for workloads with strict data residency requirements. The post explains global versus India inference profiles and walks through an example of OCR and document understanding for Indic-language documents, including structured extraction and follow-up questions in code-mixed Hinglish. It also covers prompt caching, service tiers, and CloudWatch monitoring, citing Ultrahuman's report of 80 percent lower LLM costs and 85 percent faster multi-turn chats with prompt caching.

Palette: teaching a small model to build decks in your design language

Palette fine-tunes a 20B open-weight model with LoRA to build slide decks in two roles—a designer that writes a structured JSON brief and a coder that writes pptxgenjs code per slide—trained on roughly 1,000 synthetic decks (~10,000 slides) generated by a large open teacher and filtered by a checker. On PresentBench it ranks first on Economics and Education, third on Talk, and second across the three-domain mean (59.7 vs. NotebookLM's 60.8). A check-and-repair harness that keeps a rewrite only when the checker confirms improvement adds 3.5 points on average across 131 decks.

Cities Worldwide Build and Scale AI Services With NVIDIA Physical and Agentic AI

NVIDIA and its partners are expanding physical AI and AI agents to cities worldwide, using shared computing infrastructure so city departments and local developers can build services without each acquiring their own resources. Examples include 1MillionBot's resident-service agents in Spain, LiveX's conversational AI hosts in the San Francisco Bay Area, K2K's shared vision AI infrastructure in Wrocław, and ITMAX's SamurAI V2 vision agents in Malaysia targeting 17,000+ cameras by 2027. Digital twins and local AI factories also feature, such as Honeywell's Nemotron-powered plant cognition at Dangote's Lekki refinery and Cires Technologies' sovereign AI factory at Morocco's Tanger Med port.

The latest AI news we announced in October 2026

Google recapped its October 2026 AI updates, led by the global rollout of the Gemini 4 Argon frontier reasoning model and the announcement of Gemini 4 Carbon, its strongest model yet for long-horizon engineering and cybersecurity defense, now in pre-release testing with the U.S. government and trusted cyber defenders via the Fairwind Program. The month also brought Genie 4 for real-time interactive world simulation with native audio, the Google Fitbit Edge fitness tracker, a global expansion of SynthID Detector, and a quantum computing breakthrough in which "magic state cultivation" on the Willow processor cut errors roughly 40-fold. Other items included AMIE results published in The Lancet, the Playground gaming platform, EmbeddingGemma 2, the first AI response from a Project Suncatcher satellite, and Research@ 2026 announcements.

Agentic engineering intelligence with Amazon Nova Multimodal Embeddings and Amazon Bedrock AgentCore

This post describes a multi-agent engineering assistant built on Amazon Bedrock AgentCore, Amazon Nova Multimodal Embeddings, and the open source Strands Agents SDK that assembles evidence for aerospace part substitution decisions. A planner routes queries to Catalog, Evaluation, and Knowledge specialists, with deterministic Python tolerance-rule scoring, Nova Multimodal Embeddings for multimodal retrieval and photo cosine similarity, and a supervisor that synthesizes results only when multiple specialists contribute. The engineer remains the decision-maker, and the system drafts justification memos and Integration Impact Assessments with audit trail entries aligned with AS9100 sections 8.4 and 8.5.6.

Introducing SFX v3 for developers

ElevenLabs released SFX v3, a new sound effects model in its Sound Effects API that follows multi-instruction prompts more closely than SFX v2 and handles vocals and music better, available by setting model_id to eleven_text_to_sound_v3 on the POST /v1/sound-generation endpoint. The model picks sound length from the prompt, with duration_seconds for specific timing and prompt_influence (default 0.3) to tighten prompt adherence, while looping still requires eleven_text_to_sound_v2 with loop: true. The company describes v3 as an incremental update that is less likely to clip on high-frequency sounds and recommends A/B testing common prompts across both models since they share an endpoint.

Introducing SFX v3 in ElevenCreative and ElevenAPI

ElevenLabs launched SFX v3, its newest sound effects model, which handles prompts describing multiple sounds more reliably, generates music and vocals more convincingly, and produces cleaner, more natural audio than SFX v2 in benchmark testing. It is available now in ElevenCreative (Flows, the Sound Effects tab, Studio, the Music Editor, and the mobile app) and through ElevenAPI for developers.

Accelerate agentic AI with NVIDIA Nemotron models on Amazon EKS and SageMaker HyperPod EKS

NVIDIA Nemotron 3.5 Super VL is a 120B-parameter mixture-of-experts model activating 12B parameters per token that handles text, image, document, and video input, joining other Nemotron models already on AWS. The post describes its hybrid Mamba-2, MoE, and attention architecture with a RADIO vision encoder and up to 1M-token context, then shows deployment on Amazon EKS, EKS Auto Mode, and SageMaker HyperPod EKS using NVIDIA Dynamo and vLLM, with validation and performance results on B200 and H200 instances.

Welcome Le Chonk: Mistral Large 4 on Hugging Face

Mistral Large 4 ("Le Chonk") is a multimodal, multilingual Mixture-of-Experts model with about 1T total and 49B active parameters and a 1M-token context window, released in FP8 and NVFP4 variants under the Mistral Ecosystem Open-Weight License. It combines a 72-layer text decoder using MLA with a DeepSeek Sparse Attention indexer and a 1.6B Pixtral vision encoder, and the post reports 82% on the Artificial Analysis Cyber Index vulnerability reproduce-and-patch test and 93% on Cybench. Day-0 support covers Transformers, vLLM, llama.cpp, TRL fine-tuning, Inference Endpoints and Dell Enterprise Hub, with GGUFs from Unsloth and the llama.cpp team.

Day 0 support for Mistral Large 4 on vLLM

vLLM announces day-0 support for Mistral Large 4, a 1T-parameter MoE multimodal model with 49B active parameters per token, running on the V2 model runner with breakable CUDA graphs and a new vllm/models/mistral/ package. The work adds cross-layer KV sharing, DCP/PCP context parallelism, DSpark speculative decoding, NIXL-based disaggregated serving and HiSparse KV offload, plus day-0 ROCm support on AMD Instinct MI350 Series GPUs with MXFP4 experts. Support ships in the vllm/vllm-openai:mistral-large-4 Docker image, with FP8 and NVFP4 checkpoints and serving commands documented.

Granite Switch 4.1.1 — Update models like you update software

Granite Switch 4.1.1 is a backwards-compatible patch release to Granite Switch 4.1 that adds an updated Granite Speech ASR model covering multilingual speech-to-text across 10 languages with native punctuation and capitalization, plus an updated Decision adapter in granitelib-core with improved performance on the Jev Decision Index test set. Apart from training the new ASR model and Decision adapter, nothing was recomputed and the base model is untouched. The team says it is working on tutorials and an Adapter SDK so developers and third-party providers can build their own adapters.

Fix evaluation bias to unlock better models on Amazon Bedrock

Property Finder found that 96.8% of its evaluation labels came directly from its incumbent vision model's predictions rather than human verification, biasing comparisons against alternatives. After building expert-annotated ground truth (356 UAE images) and running evolutionary prompt optimization from v0 to v27, it selected Llama 4 Maverick 17B Instruct on Amazon Bedrock, reaching a Micro F1 of 0.80 versus 0.56 for its previous model on the evaluation set, at an estimated 86% lower cost per 1,000 images.

Build specialized multimodal models with LFM3

Liquid AI introduces LFM3, a generation of Liquid Foundation Models in the 2B parameter range that places text, vision, audio, omni, and decision models on one shared backbone built from LFM3-1.7B-Base, which was pre-trained on 22T tokens and uses a tri-hybrid layer mix with a 256K context window. The release includes six 2B-class models, new in-house vision and audio encoders, and LiquidCodec, with every model fine-tunable on a single GPU and the text model fitting in about 1 GB at 4-bit precision. All LFM3 models are open-weight and available on Hugging Face, with fine-tuning notebooks and agent harness guides provided.

Genie 4: A new generation of world simulation

Google announced Genie 4, a natively multimodal world model that generates photorealistic, real-time interactive simulations with native audio, including ambient sound and interactive dialogue, and builds on Gemini models for reasoning and coherence. It is rolling out as a research preview to trusted testers, with plans to reach Google AI Ultra subscribers in Project Genie and early access partners via the Gemini API and Google AI Studio. Google notes limitations such as drift in visual details and object permanence over longer sessions, and says all generated video and audio include SynthID watermarking.