#多模态

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.

DeepMind unveils Genie 4 world model promising more realistic physics and humans

Google DeepMind has revealed Genie 4, an updated world model that improves on Genie 3 in almost every way, with more realistic physics and humans, and which paying Google customers may be able to use before too long. DeepMind frames world models as a key element toward AGI: they generate a virtual world frame by frame in real time as you explore it, without pre-rendered environments or programmed interactive elements, and could create infinite training environments for physical robots or enable rapid prototyping. The improvements are due in part to Google's updated Gemini models.

Featured coverage of this story: 《Genie 4: A new generation of world simulation》

Google’s Genie 4 world model adds audio, more realistic worlds, and real-time steering

Google DeepMind released Genie 4 in research preview, a world model that generates photorealistic, interactive worlds with real-time audio and lets users type prompts mid-simulation to change scenes, weather, camera angles, or add objects. It maintains a consistent scene for up to two minutes, double Genie 3's limit, and integrates with Gemini, which plans steps while Genie creates visuals; a Gemini Robotics experiment had a Gemini 4 Argon model break tasks into steps that Genie 4 simulated in real time. Genie 4 is available now to trusted testers and will roll out in the coming weeks to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio.

Featured coverage of this story: 《Genie 4: A new generation of world simulation》

Google’s next Genie AI model generates more realistic experiences

Google announced Genie 4, a world model for generating interactive experiences, with four main updates: more realistic worlds, real-time updates via natural language prompts, audio generation, and Gemini 3.8 Flash under the hood. It can create experiences about two minutes long, up from one minute for Genie 3, and Google pitches it for uses like filmmaker pre-visualization, game prototyping, and robot training scenarios, while warning that details and object permanence can shift over longer simulations. Genie 4 is launching first as a research preview to trusted testers, will be "coming soon" to Google AI Ultra subscribers, and Google has opened a waitlist to try it in Project Genie.

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.

Show HN: EveryGen AI – AI images, short videos and copy rewrites in one tab

EveryGen AI is a browser-based tool that combines image generation (defaulting to GPT Image 2), video generation (defaulting to MiniMax H3 Max Turbo), browser image utilities, and a Gemini 2.5 Flash text-rewriting tool called Humanize AI. It offers annual credit plans from Lite ($179/year for 7,200 credits) to Max ($1,199/year for 96,000 credits), and states it is an independent tool, not an official MiniMax or fal website.

Can You Make a 3D Game with AI, If You Don't Know How to Do Anything?

The author describes building a browser-based 3D game intro inspired by Mass Effect using mostly free tools plus Codex from a $20 ChatGPT subscription, spending about $15 and roughly 10 days. The workflow combined ChatGPT for prompts and character images, Hunyuan3D for models, Mixamo for animation, and Freesound for audio, with the author rewriting dialogues and fixing bugs himself. He concludes that AI now lets beginners experiment with game development without deep technical knowledge.

Can AI draw San Francisco from memory?

VOYGR tested six chat models from Anthropic, OpenAI and Google on 41,526 San Francisco coordinates, asking what is at each point from memory with no maps or internet, and checked every answer against OpenStreetMap. At a point every 100 m, Gemini 4 Argon, GPT-6.1 Sol and GPT-6 Astra tied at 86% correct on water, park or everything else, against 61% for answering "not water, not park" everywhere, while on a 500 x 500 m Marina block the best street score was 0.32 versus 0.23 for shuffled answers. The authors conclude models capture the coastline and large parks but not street grids, and recommend a geocoder rather than model memory for reverse geocoding.

Apple Creator Studio adds powerful new AI features and even more creative tools

Apple announced updates to Apple Creator Studio, including built-in MotionVFX and Content Hub stock video in Final Cut Pro for Mac, an intelligent assistant in beta and an enhanced Stem Splitter in Logic Pro for Mac, and Generate Fill in Pixelmator Pro. Keynote, Pages, and Numbers gain AI assistants in beta that understand document context plus Model Context Protocol support for workflow automation. The updates arrive on the App Store this November, free for existing subscribers, with new subscriptions at $12.99 per month or $129 per year.

vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model

The vLLM Semantic Router team released d3-ultra, a 26.09B-parameter dense multimodal decision model under Apache-2.0 that does not generate text but reads an input once and returns probabilities for typed questions (choice, yes/no, score, box, track), with 131,072 tokens of context per question. It scored 87.2 on the team's private SR-Text test, the highest of six decision models compared, but trailed its sibling d3-turbo on SR-Box grounding (58.2% vs 60.0% over 10 domains). It was tested on one AMD Instinct MI325X with BF16 weights, where 255 questions about a photo took 2.16 s, and is available via Transformers, the vLLM Semantic Router Engine, a free API and a Hugging Face Space.

Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing

Baidu released Unlimited OCR 2, a 3.34B-parameter OCR model under the MIT license that parses documents of hundreds of pages in a single request using SR-SWA, a sparse attention scheme that keeps per-step decoding cost nearly flat. Baidu reports first place on MPDocBench (84.6) and MarathonDocBench (77.1), with heading hierarchy its lowest sub-score at 48.2 on long documents. All benchmark figures are self-reported, and the model runs on one CUDA GPU in BF16 with code on GitHub.

OpenAI Launches GPT-6.1 Terra, Halving GPT-6.1 Sol’s API Price for Agent Work

OpenAI released GPT-6.1 Terra, a closed-weights model for visual reasoning, computer use and writing with a 1.05M-token context window and 128K max output, priced at $1 per million input tokens and $5 per million output tokens—half of GPT-6.1 Sol. OpenAI reports 85.5% on MMMU-Pro (Vision) at max effort, 60.0% on Agents' Last Exam v1 at $1.01 per task, and 50.3% on Terminal-Bench 4.0, trailing GPT-6.1 Sol (55.9%) and Claude Sonnet 5.5 (70.6%). It is available in the API as gpt-6.1-terra and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.

Featured coverage of this story: 《Introducing GPT-6.1 Terra》

GPT-6.1 Terra

A link-blog post notes that OpenAI's GPT-6.1 Terra announcement emphasized visual reasoning and computer use, describing it as a fast, efficient model positioned between the cheaper Luna and the more expensive Sol, priced at $1/million input tokens and $5/million output. The post quotes OpenAI's Ted Sanders saying Terra is better than Astra and Sol at technical communication, similar on vision and computer use but faster, and a bit worse on most other things, and includes pelican test images across reasoning levels.

Featured coverage of this story: 《Introducing GPT-6.1 Terra》

Blue Planet III is keeping the rough edges in the age of AI

BBC's Blue Planet III deliberately leaves in imperfect shots, such as plankton in the water or visible boats and buoys, to counter the overly polished look associated with AI. Executive Producer Elizabeth White told The Guardian that shots previously cut for such flaws are now kept to present a more authentic, organic view of the oceans. The series features dolphins trapping fish with sea sponges, inter-species fish communication, and 200,000 crabs lining up in formation to feed.

OlmoEarth v1.3: State-of-the-art embeddings for the whole planet

OlmoEarth v1.3 is a new satellite embedding model that reads a year of Sentinel-1, Sentinel-2 and Landsat time series to produce one 128-dimensional vector per 10-meter pixel, with the first 64 dimensions usable as a smaller embedding. Precomputed annual global embeddings are shared on Source Cooperative, and a browser-based OlmoEarth Embeddings Explorer lets users explore maps, find similar places, and train small classifiers without a GPU or sign-in. The 128-dimensional embeddings scored highest on 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, and the model took 53% of v1.2's computing time and 18% of v1's, with weights, pretraining data and code all open.

NVIDIA Transforms European Retail, From Digital Discovery to Physical Operations

European retailers are using NVIDIA-accelerated open models and physical AI across commerce and fulfillment, including Shopify post-training NVIDIA Nemotron 3.5 Lightning for its Sidekick GraphQL subagent, which it says surpasses a leading frontier model in accuracy on an internal benchmark while cutting serving costs by 90% and delivering responses 7x faster. Otto Group extended its autonomous warehouse initiative, built with Reply on NVIDIA Omniverse and running on Google Cloud, to its largest fulfillment center in Haldensleben, Germany, and Nebius is showcasing three early-access NVIDIA retail blueprints at GTC Berlin. Other cited deployments include Zooplus with Infosys, Photoroom with Vestiaire Collective, L'Oréal with Omi and NVIDIA ALCHEMI, Mirakl, ALLSIDES, and CATCHES with AMIRI.

Empowering frontline maternal care with AI-assisted ultrasound

Google researchers evaluated an AI model that estimates gestational age from standardized "blind sweep" ultrasounds performed by novice operators, reporting a mean absolute error of 4.18 days versus 4.48 days for expert sonographers across cohorts in Chicago and Nairobi. The model, minimally tuned on 120 Chicago participants and tested with no Nairobi data, achieved MAEs of 4.10 and 4.27 days respectively, as published in JAMA Network Open.

Research@ 2026: 10 ways we're amplifying human ingenuity

At its Research@ event in Mountain View, Google Research outlined 10 updates, including expanded early access to Google Earth AI's experimental geospatial capabilities, a multi-agent orchestrator for reproducing prior methods, and public access to its Deep Review verification system. Other highlights include the Synaptics Astra SR200 Series chip with Google's co-designed Coral NPU, upcoming integration of TimesFM and TabFM into Gemini for business, and a Lancet study in which 98 patients consulted AMIE ahead of urgent care visits. Google.org is also providing $10 million in funding for crisis resilience and public health work.

Gemini 4 Argon: Built to solve your hardest problems

Google announced Gemini 4 Argon, a frontier model now available globally via the Gemini API in AI Studio and the Gemini platform, Google Antigravity, the Gemini app, and AI Mode in Search, priced at $2 per million input tokens and $10 per million output tokens. Google reports Argon ranked #1 on the Vals Index and cites internal and partner uses including a Chrome architectural refactor, a 2.85x inference speedup on Gemma 4 E2B, quantum computing and fusion research, rare disease analysis, enzyme design, and a robotics video-to-sim-to-real pipeline. Google also described safety work including jailbreak defenses, agent misalignment monitoring, and a 0.7% attack success rate on the Gray Swan indirect prompt injection benchmark.

Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space

This Import AI issue covers the Humanity's Sixth Sense benchmark from Elorian and Scale AI, where humans score 93.1 pass@1 versus 53.6 for the best model, GPT 6 Astra, with Chinese open weight models scoring far lower. It also discusses P-Zero Research's TasteVal benchmark, which reports that frontier models' experimental research taste has doubled every 3.0 months since December 2025, with Opus 5.5 exceeding the human baseline, and covers Reflection's 501B-parameter Beam and Mistral Large 4, both trailing Chinese open weight models on several benchmarks.

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Google DeepMind announced EmbeddingGemma 2, a 740-million-parameter model built on the Gemma 4 architecture and released under the Apache 2.0 license, which maps text, code, images, video, and audio into a unified embedding space for on-device use. It offers modular encoders (as little as 270M parameters for text-only, with optional 170M vision and 300M audio encoders), Matryoshka Representation Learning to truncate vectors from 768 down to 128 dimensions, an 8K token context window, and on a Google Pixel 11 Pro requires as little as ~191MB active RAM for text-only weights and ~567MB for the full multimodal model. The company reports a 9.92-point improvement on MTEB Code (68.76 to 78.68) over EmbeddingGemma and says the model is available on Hugging Face and Kaggle, with support across tools including transformers, vLLM, llama.cpp, Ollama, and Qdrant.

Introducing Mistral Large 4

Mistral launched a public preview of Mistral Large 4, a 1-trillion-parameter natively multimodal model with 52 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, with weights to be released by the end of the month. Mistral reports competitive performance across coding, agentic workflows, multimodal understanding, cybersecurity, science and knowledge work, claiming it outperforms any open-weight model developed in the US or Europe. The preview API is available on Mistral Studio, and the model will serve as the foundation for a new generation of specialized Mistral models.