Google DeepMind released EmbeddingGemma 2, a 740M-parameter model built on the Gemma 4 architecture under the Apache 2.0 license, mapping text, code, images, video and audio into a shared embedding space.
Why it matters: The official release provides concrete metrics such as parameter count, memory footprint and benchmark scores, helping readers judge whether it suits on-device retrieval scenarios.
AI ad creative platform Melius announced on Tuesday a total of $25 million in funding, including a $20 million Series A led by CRV and a $5 million seed round led by General Catalyst. Founded by three former Ramp engineers, the company pivoted to a "creative work agent lab" that generates ad assets and campaigns after abandoning its original AI marketing tool and rewriting its entire codebase, and says it topped $1 million in annualized revenue within two months of stealth exit in July.
Microsoft AI released MAI-Transcribe-2-Streaming, a real-time speech transcription model it says ranks first in accuracy on Artificial Analysis, supporting 60 languages, with first partial results delayed by just over 100 milliseconds and pricing of $0.54 per hour of audio before the end of the year, alongside the voice model MAI-Voice-2.1.
Google today launched Guided Vision in Gemini Live on compatible Android devices, providing real-time voice descriptions of what the phone camera sees, helping users read fine print, identify objects, and describe their surroundings.
OpenAI announced the global rollout of two new ChatGPT shopping features: virtual try-on and Favorites. Users can upload selfies or full-body photos to try on clothing and accessories, or upload product screenshots to request try-ons; the try-on feature is based on the newly released ChatGPT Images 2.5 model. Favorites lets users save products to the in-app Library.
Tavus launched Griffin, which it calls the first Human Interaction Model, capable of processing speech, facial expressions, tone, gestures and pauses in real time and generating video bidirectionally. In the company's research, 48% of participants thought Griffin was a real person after a one-minute video call, versus a previous high of 2%; in what Tavus says was independent testing by Nvidia, Griffin scored 3.83, close to 3.92 for a real person.
TechCrunch interviewed Airbnb CEO Brian Chesky, who, as the company launched AI search this week, argued that AI agents need operating-system-level support, saying iOS, macOS, and Windows are not AI operating systems and that AI should ideally run at the kernel level.
Condé Nast partnered with the AWS Generative AI Innovation Center to build a multimodal video discovery solution based on Amazon Bedrock and Amazon OpenSearch Service, using the TwelveLabs Marengo embedding model to jointly encode visual, audio, and transcript signals.
In its fall product update, Airbnb introduced AI search, letting users search listings via text or voice prompts with a toggle; the interface shows dynamic filters based on preferences—for example, typing baby automatically surfaces filters for cribs and playgrounds.
Google released its new model Gemini 4 Argon, saying it excels at cybersecurity, coding, and research, and can autonomously discover, verify, and fix critical software vulnerabilities. The model was specifically trained on defensive cybersecurity tasks and is currently available only to select cybersecurity partners through the Fairwind security program; Google, citing Vals' model index, says it leads OpenAI GPT-6 Astra and Anthropic's Fable and Opus on multiple benchmarks.
Hugging Face has launched the Open TTS Leaderboard, evaluating open-source TTS models with objective metrics including Qwen3 ASR WER/CER, RTFx and TTFA on H200, and WavLM speaker similarity, cutting per-model evaluation time from about two weeks to hours.
Why it matters: The post explains how objective metrics shrink TTS evaluation from weeks to hours, and why mainstream arenas structurally underrepresent open-source models.
Microsoft Research launched Quine, a research system combining a multimodal world model of biology with an interactive framework connecting scientific tools, literature, and researchers. With the Broad Institute, it applied the system to pancreatic ductal adenocarcinoma (PDAC) research, predicting and ranking compounds that could drive tumor cell state transitions; wet-lab validation confirmed top-ranked candidates, and narrowing the compound pool to validation candidates took just one weekend.
Why it matters: Microsoft's official introduction covers Quine's architecture and real PDAC cancer research results, showing readers how AI world models can compress wet-lab iteration cycles.
AMD acquired World Labs; since AMD is a listed company, the author notes the acquisition price was $8.2 billion. World Labs, founded in 2024, built training teams for image, video and spatial reconstruction, and after acquiring SceniX targeted robot simulation. Its Atlas model, trained from scratch, can predict the next viewpoint from 2D images, combining generative models with multi-view geometry and essentially solving the long-standing sparse reconstruction problem in computer vision, with applications in robotics RL environments, scene generation and real estate reconstruction.
Microsoft Research Asia – Singapore (MSRA – Singapore), which opened on July 24, 2025, is Microsoft's first research lab in Southeast Asia.
AWS demonstrates an agent-driven synthetic monitoring solution based on Amazon Nova Act and Amazon Bedrock AgentCore, replacing traditional DOM-selector-based scripts with natural language actions and reducing maintenance costs caused by UI changes.
AWS showed how to use the open-source RL framework SkyRL on SageMaker HyperPod to train the Qwen3-VL-8B vision-language model on a visual maze navigation task. Starting from a VisGym SFT checkpoint, GRPO post-training raised the solve rate on a fixed set of 64 mazes from 43.75% to over 95%.
Runway has released the research preview of GWM Worlds 2, which uses the WorldPrompt format to control characters, cameras, and environments in simulated worlds via prompts, generating 720p, 24 fps video and 48,000 Hz audio in real time.
Google Research unveiled its AI video co-director research, building a multi-agent orchestration layer based on Gemini and Veo with four frameworks — Co-Director, CANVAS, A²RD and VQQA — that reformulates long-form video generation as a global optimization and world-state tracking problem.
Google DeepMind has launched Gemini 3.8 Live with Live Avatar, combining near-real-time video generation with speech to give Gemini's conversational AI a dynamic visual presence, with accurate lip sync, natural expressions and smooth turn-taking.
Liquid AI has released experimental DSpark draft models for the vision-language model LFM2.5-VL-3B, using speculative decoding to trade for faster inference without changing output quality — decoding speeds up to 3.13x on-device and 2.66x on H100, with end-to-end gains up to 2.62x and 2.27x respectively.
Why it matters: The article details a speculative decoding acceleration scheme for vision-language models with concrete speedups and memory overhead, plus integration guidance for llama.cpp, MLX-VLM, and SGLang.
invideo uses GPT-6 Astra to boost video editing: more precise edit planning, 3x faster color correction and grading, and up to 50 custom effects produced in a day.
NVIDIA held AI Day Singapore on September 22-23 at the Raffles City Convention Centre, where it and its partners showcased AI advancements across the Southeast Asia region.
Google Research launched a new research experiment where teachers can use generative UI (GenUI) to dynamically generate customized, guided interactive learning simulations based on their teaching goals and curriculum. Built on active learning principles from the LearnLM education model family, the release includes over 30 AI-generated, teacher-reviewed English STEM learning interactives covering physics, chemistry, biology and math for middle and high school.
NVIDIA Vera Rubin NVL72 made its debut in MLPerf Inference v6.1 preview submissions, delivering up to 3.7x the throughput of GB300 NVL72 on Qwen3-VL and 2.5x on DeepSeek-R1.
The University of Manchester, working with the NVIDIA Earth-2 team, repurposed the Earth-2 CorrDiff generative downscaling model for air pollution forecasting and added Earth-2 StormCast, which can directly leverage air quality observations.
H Company released NeoMME, a multilingual multimodal encoder in 260M and 800M sizes. A single bidirectional Transformer processes both text tokens and 32×32 image patches, trained from scratch with a masked discrete diffusion objective, with no reliance on a pretrained vision tower or a causal language model. Both versions are open-sourced under Apache 2.0.
Google DeepMind released agentic video understanding for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite, replacing static 1 FPS processing with dynamic searching and inspection of video segments; official benchmarks show up to 66% lower cost, up to 88% fewer tokens, and up to 7% higher accuracy.
Why it matters: The official release gives concrete cost, token and accuracy figures plus how to enable it, letting developers directly assess switching their existing video processing pipelines.
Microsoft Research released GigaPath-Flash and GigaTIME-Flash, using knowledge distillation to compress the billion-parameter GigaPath encoder into a 22M-parameter ViT-S encoder, with performance within 3% of the original model at roughly 50x lower compute.
Google Earth AI has introduced an experimental research capability called the Planetary Prediction Engine (PPE), which autonomously performs the full geospatial prediction workflow—data discovery, cleaning, feature engineering, model training, and evaluation—starting from natural language queries, compressing work that previously took weeks of manual data engineering down to a few minutes.
On August 27, Google DeepMind released Gemini Omni 1.1 Flash, bringing a suite of creative controls for generative video aimed at developers, available via the Gemini API in Google AI Studio.
Why it matters: The official release provides specific parameters for new capabilities like scene extension, keyframe interpolation and 4K upscaling, along with access channels, so developers can evaluate adopting it in generative video workflows.
Google is presenting AgentHands at CHI 2026, a research prototype that lets AI agents generate speech-synced hand gestures in immersive environments like Android XR, building on Project Astra and Gemini 3.1 Flash Live.
Google Research introduces the Mobility-Embedded POIs (ME-POIs) framework, which folds aggregated, anonymized visit-mobility patterns into language models' place representations, capturing POIs' temporal activity rhythms.
Microsoft Research released Skala 1.1, trained on 2.5 times the data of the previous version, ranking first in 32 of the 55 categories of the GMTKN55 benchmark with a weighted average error of 2.8 kcal/mol at compute cost comparable to a meta-GGA functional. Skala is now available in CP2K and being integrated into Psi4, FHI-aims, ORCA, and VASP, alongside a dynamic benchmark tracking each version's computational performance.
Google Research has introduced PhotoScan, a deep learning framework that estimates body fat percentage, Android-to-Gynoid fat ratio, and visceral-to-subcutaneous fat ratio from standard 2D smartphone photos to predict insulin resistance.
Hugging Face's OlmoEarth Studio now supports computing and exporting embeddings from the open-source OlmoEarth foundation model, output in Cloud-Optimized GeoTIFF (COG) format.
Microsoft Research introduced the MindTopo benchmark, evaluating multimodal large models' ability to reason about and plan with topological concepts such as connectivity, closure, ordering, separation, and knots. Testing shows models perform markedly better on static image recognition than on interactive planning tasks, and both fall far short of human level, with errors mostly involving lost structural relations during planning or physically infeasible actions. The benchmark aims to provide a controlled diagnostic tool for research on agents in robotics and interactive environments.
Google DeepMind released a large-scale multilingual sign language-to-text (SL2T) translation model, bringing new sign language input features to deaf and hard-of-hearing users, first supporting American Sign Language (ASL) to English in Gboard and Live Transcribe on Pixel 11, with more devices and languages to follow.
Why it matters: The article gives concrete details on training scale, benchmark scores, and privacy design, letting readers understand the real technical path to deploying sign language translation.
Google Research has launched AMIE (Video), built on Gemini and Project Astra, which can conduct real-time synchronized clinical video consultations, perceive non-verbal cues, and guide virtual physical exams.
Why it matters: The article covers the architecture, evaluation scale, and quantitative comparisons against primary care physicians (PCPs), while candidly noting limitations like simulated scenarios, helping readers understand the practical boundaries of audio-visual clinical AI.
Microsoft Research introduces CARE-X, a research model and unified chest X-ray VLM built on SigLIP2-so400M and Phi-4-mini-instruct (3.8B), jointly trained with auxiliary heads to provide report generation and structured predictions with confidence, and aligned to clinical rewards using DAPO reinforcement learning.
Meta today released Muse Glimmer, a 30B-parameter open-source multimodal model distilled from Muse under the Apache 2.0 license, aimed at privacy-focused local deployment and agentic uses like coding, document analysis, and personal assistance.
Why it matters: The article provides architecture details, agent benchmark comparisons, and day-0 usage across inference stacks, helping readers evaluate local multimodal deployment options.