Note that most of the examples are visual. See my Youtube video for the best experience.

Qwen3.8-Flash-Next previews the architecture Qwen says will underpin Qwen4 while focusing on lower inference cost. Its hybrid Gated DeltaNet and sparse attention design supports 262K tokens natively, uses only 6B activated parameters from a 125B model, and is being released as open weights with a managed Qwen3.8-Flash API planned. Read more

Tencent’s Hy4 preview is a 770B-parameter open Mixture-of-Experts model with 49B active parameters and a 1M-token context window. The Hugging Face release targets hard coding, office, game, and research tasks and includes vLLM and SGLang deployment recipes under Apache 2.0. Read more

Gemini 3.5 Transcribe is Google’s new speech-to-text model for cleaner, more capable voice applications. It supports low-latency streaming, speaker labels and word timestamps for recordings, custom vocabulary, more than 85 languages, and automatic cleanup of filler words and self-corrections. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

Skild’s S1 robot learns a new manipulation task from one video demonstration instead of task-specific fine-tuning. The company reports that it can handle unseen tasks lasting up to 10 minutes, with one example roughly matching 380 post-training demonstrations in its experiments. Read more

FastVideo and partners released FastH3 Preview v1, an open-weight text-to-video-and-audio model distilled from MiniMax H3. The recommended four-step checkpoint can generate 15-second 768p clips in under 13 seconds on eight B200 GPUs and is built to support multiple resolutions, aspect ratios, and durations. Read more

BRIA’s Fibo 1.5 makes structured, controllable image generation much faster through few-step distillation. The 8B model keeps FIBO’s JSON-native prompting and licensed-data approach while generating in roughly 4–6 steps without classifier-free guidance. Read more

RSI-Exam turns recursive self-improvement into a test of whether an AI agent can repeatedly improve an executable method. Its first release contains 88 tasks across six domains, uses hidden-set re-execution to check whether gains generalize, and records the experiments and discarded ideas along the way. Read more

GLM is Z.AI’s flagship open-source model built for long, difficult engineering tasks like coding, debugging, and autonomous agent work. It delivers top-tier performance while being dramatically cheaper and faster than frontier closed models. Get 10% OFF HERE.

Google’s Gemini Omni 1.1 Flash gives developers tighter control over AI-generated video. It can extend scenes up to 40 seconds, interpolate between first and last frames, create faster 360p drafts, accept short video references, and upscale results to 4K through Google’s developer platforms. Read more

Google’s Planetary Prediction Engine is an experimental agent that builds geospatial prediction models from natural-language questions. It discovers and cleans data, engineers features, trains and evaluates models, and generates a report, with Google saying the workflow can shrink from weeks of manual work to minutes. Read more

Zero-WAM teaches a robot to use human videos as in-context task instructions when language alone is too vague. Its authors report 74.2K human–robot pairs across 8.6K tasks and a 46.95% average success rate on seven unseen RoboTwin 2.0 tasks, 29.5 points above their video-action baseline. Read more

Code World Model uses a coding agent as the “brain” that maintains persistent world state while a video model renders what happens. Instead of asking a video generator to remember rules by itself, the system updates executable state and converts it into proxy videos that guide consistent visual evolution. Read more

OVOW turns one monocular video into separate, simulation-ready 4D mesh objects with recovered motion. Its training-free pipeline combines vision-language scene decomposition, per-instance reconstruction, pose and deformation recovery, and physics-grounded assembly for downstream simulation and editing. Read more

FixAnything uses a pretrained video diffusion model to repair 3D rendering artifacts while preserving the camera path. It works across 3D Gaussian splats, NeRFs, meshes, and sparse point clouds, and its geometry-aware preference stage improved pose accuracy from 61.1 to 68.3 AUC@5° in the authors’ tests. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like “um,” fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

DiffusionOPSD replaces a single end-of-training image reward with explicit positive and negative targets during diffusion training. The authors report stronger held-out results than competing methods while cutting profiled training GPU-hours by 40% on SD3.5-M and 63% on Z-Image-Turbo. Read more

Block3D speeds up text-to-3D generation by denoising shape-token blocks in parallel while correcting low-confidence tokens before they are finalized. On a held-out TRELLIS-500K set, it reduced mean end-to-end generation time from 25.71 seconds to 4.99 seconds, a 5.15× speedup without sacrificing geometric fidelity. Read more

VoiceMem gives voice assistants two kinds of memory: factual recall on the left and emotion, preference, and personality on the right. Its streaming pipeline retrieves memories while a person is still speaking, and the project reports 91.2% on LoCoMo with 430 memory tokens and 134 ms latency in its benchmark setup. Read more

ORBIT builds a harder structure-from-motion benchmark from real 360-degree web videos. The project uses panoramic trajectories to render difficult perspective clips containing blur, motion, crowds, and low texture, exposing weaknesses in both classical and recent camera-pose methods. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

Evoke is an open-source AI world model that can generate an interactive world, remember what happened, and keep it running for hours. You can move around it like a game and change the world with new instructions while it is already being generated. Read more

LLM-as-a-Verifier uses an AI model as a judge that gives detailed feedback on whether another AI is doing a task correctly. Instead of simply saying “pass” or “fail,” it can track progress and provide signals that help AI systems improve in areas like coding, robotics, and medicine. Read more

GEN-1.5 is a robot AI model that can learn some new physical tasks after seeing just one example, without needing extra training. This means you could potentially show a robot how to do something and have it quickly copy the skill in new situations. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

SenseNova U1.5 8B MoT is an 8-billion-parameter model presented with a mixture-of-tokens design. At this scale, the model is interesting to developers who want something more manageable than a huge frontier system. Smaller open models can be easier to run locally, fine-tune, and embed into products, although real performance still needs independent testing. Read more

Ornith-1.5 is an open AI model that tries to improve itself by inventing new problems, figuring out strategies, and learning from its own attempts. It comes in sizes ranging from a huge 397B-parameter model to a small version designed to run on phones. Read more

Ling-3.0 Flash Base is a highly efficient open AI model with 124 billion total parameters but only activates about 5.1 billion at a time. Its sparse design and hybrid attention system are built to handle long inputs efficiently while keeping computing costs lower. Read more

GLM is Z.AI's flagship open-source model built for long, difficult engineering tasks like coding, debugging, and autonomous agent work. It delivers top-tier performance while being dramatically cheaper and faster than frontier closed models. Get 10% OFF HERE.

GeoWeaver can turn long videos into a consistent 3D reconstruction without losing track of where the camera and objects are over time. It breaks the video into manageable pieces and then intelligently stitches them together to reduce drifting and misalignment. Read more

HappyShrimp 1.0 is Alibaba’s AI music model that can turn a simple description of a mood, story, or musical style into a complete song. It generates the lyrics, melody, arrangement, and vocals together, so you do not need to know music theory to use it. Read more

HarnessEval is a new way to test AI world models by using an AI “harness” that investigates the output like a detective instead of just giving it a simple score. It can break an evaluation into smaller tasks, use specialized sub-agents and tools, and check things like realism, physical consistency, and whether actions actually work as expected. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like "um," fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

Audio8 TTS Preview is a very small AI voice model that can generate speech and clone a voice from an example without requiring a huge model. It supports English and Chinese plus several experimental languages, aiming to make high-quality text-to-speech easier to run on modest hardware

Qwen-Video-Edit turns an image-editing AI into a video editor, allowing you to change videos using simple text instructions. Instead of training a giant video model from scratch, it adapts an existing image editor to work directly with video data and can apply different edits across longer videos. Read more

4DAnyone can turn a normal single-camera video of a person into a 4D version that can be viewed from new angles over time. It does this without special camera setups, motion-capture suits, or calibration, making it possible to reconstruct moving people from casual videos. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

NVIDIA’s AVO architecture reached a perfect score on ARC-AGI-3, a benchmark built around difficult visual and interactive tasks. The significance is less about one leaderboard number and more about the kind of work being tested: agents must interpret changing environments, plan across many steps, and act reliably instead of answering a single isolated prompt. Read more

Anthropic showed that Claude can help scientists design proteins and analyze chemistry much faster than traditional workflows. In one test, it successfully designed protein binders for 14 out of 15 targets and also completed chemistry analysis in minutes using raw lab data. Read more

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

Google introduced Gemini 3.7 Flash as its new workhorse model for coding and agents. It is positioned for everyday, high-volume tasks where developers need capable AI without the latency cost of a larger model. Read more

xAI’s Grok 4.6 focuses on long-running agents and more ambitious interactive and visual work. The release builds on Grok 4.5 and signals xAI’s push toward assistants that can stay useful across bigger, more involved tasks. Read more

DeepSeek has published DeepSeek-V4-Pro-0813 on Hugging Face. The release makes a new Pro checkpoint available for builders to inspect and experiment with through the open-model ecosystem. Read more

Qwen released Qwen3.8-2.4T-A95B, a very large mixture-of-experts model with 95 billion active parameters. Its Hugging Face release gives researchers and infrastructure teams another massive open model to evaluate and deploy. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

Qwen also released Qwen3.8-27B, a more compact member of the same model family. A model at this size is far more practical for teams that want strong capabilities without the hardware demands of a trillion-scale system. Read more

Z.AI’s GLM-5.3 is the latest iteration in its line of models built for difficult engineering and agent work. It continues the company’s emphasis on models that can plan, debug, and carry out longer coding workflows. Read more

NVIDIA released Nemotron 3.5 Lightning alongside NeMo Switchyard for faster, more efficient agentic AI. The lightweight open model and routing library are meant to give teams more control over models, data, and workflows from PCs to data centers. Read more

Meta’s Muse Glimmer is an open 30B model designed for always-on local agents. It can run on a single GPU and is tuned for tool use, long tasks, and recovering from failures, under an Apache 2.0 license. Read more

Cactus released Needle 2, a 45M-parameter agentic model that ships as a 14 MB binary. It is designed for tool calling, device use, and structured extraction on tiny devices, using only 28 MB of session memory. Read more

GLM is Z.AI's flagship open-source model built for long, difficult engineering tasks like coding, debugging, and autonomous agent work. It delivers top-tier performance while being dramatically cheaper and faster than frontier closed models. Get 10% OFF HERE.

LTX-2.5 is an open-weights video foundation model for generating multi-shot scenes in one pass. It can edit real footage, supports high-end EXR output, and can be fine-tuned or run on your own hardware. Read more

Sand.ai previewed Magi-2, its next video-generation model focused on scaling efficiently. The company is working toward video systems that learn how objects, action, space, time, and sound fit together rather than simply producing convincing frames. Read more

Tencent’s WorldClaw turns one open-ended prompt into an explorable, editable open-world 3D scene. The agentic framework is aimed at generating 3D worlds at scale, making it easier to create environments that can be explored and changed. Read more

MiniMax Music 3.0 is a new open-weights music model built for production use. Given a concept and optional lyrics, it can compose, arrange, perform, and produce a complete song. Read more

IndexTTS 2.5 improves zero-shot speech generation across more languages, with faster inference and stronger quality. The model supports Chinese, English, Japanese, Spanish, and Arabic while transferring emotion to languages it has not explicitly seen in training. Read more

MiDashengLM-Gen is a system for generating complete audio scenes from structured text descriptions. It can blend speech, music, sound effects, and environmental audio into coherent clips, while supporting nine languages and emotion control. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like “um,” fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

SCoPE gives video-generation models a better way to understand and control a camera’s viewpoint. By giving each video token its camera ray as a second position signal, it promises more precise camera motion in existing video diffusion models. Read more

FlaRe combines editable, ray-traceable geometry with neural radiance fields for photorealistic scene rendering. Its representation is built for interactive rendering plus useful operations such as deformation, mesh extraction, and stylization. Read more

Google DeepMind is putting sign-language AI directly into users’ hands. The project brings accessibility-focused AI closer to practical daily use for people who communicate through sign language. Read more

MatrAIx is open-source infrastructure for testing products and AI systems with simulated users before they reach the public. Its platform models a population of persona agents and can evaluate surveys, chatbots, web prototypes, and apps. Read more

A new survey organizes reliable self-evolving agents into five levels of increasing autonomy. It pairs each type of agent self-improvement with evidence and controls, offering a clearer way to think about systems that change themselves. Read more

Dots 3 is a new AI release from Dots, presented through its English-language studio page. The project adds another model effort to watch from the rapidly expanding Chinese AI ecosystem. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

Google DeepMind open-sourced WeatherNext models that improve cyclone forecasting. The team says WeatherNext Cyclones gives forecasters about an extra day of useful warning for storm track, intensity, and wind structure, which can matter when cities are preparing for hurricanes and typhoons. Read more

Alibaba released Qwen3.8-Max, its strongest Qwen model so far. The model is aimed at coding and long agentic work, and Alibaba says it will return to an open-weight release strategy so developers can run and adapt a frontier-scale Chinese model. Read more

Meta is positioning Muse Spark as a serious coding and agentic AI model for developers. The latest Muse Spark 1.2 highlights faster coding workflows, multimodal reasoning, long context, and lower-latency agent work through Meta's model API. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

Stanford scientists used a DNA-trained AI model to design 16 working bacteria-infecting viruses. The study points to useful phage therapies for antibiotic-resistant bacteria, but it also raises biosecurity questions because AI can now help design functional genetic systems not found in nature. Read more

Xiaomi open-sourced Xiaomi-Robotics-1, a large robot foundation model. It was trained on more than 100,000 hours of real-world manipulation trajectories and is designed to help robots handle unseen environments, new objects, and new tasks with only a small amount of extra demonstration data. Read more

Tencent Hunyuan introduced Hunyuan3D-Buffalo 1.0 for 3D generation, understanding, and editing. The system connects language, 3D representations, and generative modules so one model can answer questions about 3D objects, generate assets from prompts, edit meshes, and break objects into parts. Read more

Alibaba's Tongyi Lab showed Wan-Animate-2 for real-time character animation. It can drive a character directly from reference video, preserve identity better, control camera viewpoint with text, and has a Lite version aimed at live avatars and streaming hosts. Read more

Run the best AI video model, MiniMax H3, with a flexible pay-as-you-go subscription. Upgrade your videos from 768P to sharp 2K quality for just $0.05 per second, and use Prompt Expansion to make your prompts more accurate, cut down on trial-and-error, and get better results at $3.60 per M tokens. Try it today!

BigBang is a new model project built around self-evolving synthetic tasks. The team says it post-trained a Qwen3.6-35B-A3B model using a generator-critic loop that creates verifiable frontier research tasks, reaching performance between much larger DeepSeek V4 systems. Read more

LongHorizon-Harness is a new framework for agents that need to work across long tasks. Instead of letting one agent keep an ever-growing memory, it splits work into manager, executor, and auditor roles so only independently checked facts guide the next step. Read more

Multi-Agent-CAD turns text prompts into CAD models with a structured multi-agent pipeline. The project splits planning, geometry design, code generation, and repair into separate stages, cutting token use sharply while keeping strong feature pass rates on mechanical-part benchmarks. Read more

Alibaba DAMO released ClinFusion for medical multimodal AI. It combines 2D and 3D medical image understanding, report generation, instruction following, and text medical tasks, with open 8B and 32B checkpoints for researchers to test. Read more

Rei Labs introduced Adapt-1 Preview as a test-time learning research system. Instead of relying on a fixed pretrained model, Adapt-1 is designed to update from scored feedback while preserving task rules, evidence provenance, and inspectable decision state. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like "um," fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

LeapTalk is a real-time talking-head generation method that claims up to 200 FPS on one H200 GPU. It distills a diffusion-style teacher into a one-step streaming model, trying to keep identity stable and lips synced over long videos. Read more

SymphonyGen is a research system for controllable orchestral music generation. It uses a harmony skeleton like a short score, then generates full multi-track orchestration while trying to reduce unwanted dissonance and preserve musical structure. Read more

VocalRender turns composer-style scores directly into expressive singing audio. Instead of requiring hand-made phoneme timing, it reads lyrics, MIDI pitches, note values, and tempo, then generates singing with pitch, timbre, and prosody details. Read more

TRACE released code for active scene reconstruction in robotics. Its planner tries to move a camera toward useful surface regions instead of repeatedly scanning the same geometry, helping robots build better 3D maps under a time budget. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

Google DeepMind introduced Gemini Robotics 2, a set of models that help robots reason, move, and work together in the real world. It can control full humanoids from feet to fingertips, run on-device, and adapt to new robot bodies with only a few hours of data. Read more

OpenAI released Codex Security as a CLI and TypeScript SDK for finding, validating, and fixing security bugs in code. Developers can run scans locally or in CI, compare scan histories, and use either ChatGPT sign-in or API-key based authentication. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

Thinking Machines Lab released Inkling-Small, an open-weights multimodal reasoning model that is much smaller than its flagship Inkling model. It has 276B total parameters with 12B active, supports images and audio, reaches up to a 1M-token context window, and is available through Hugging Face and Tinker. Read more

AMD introduced Instella-MoE, a fully open mixture-of-experts language model trained on AMD Instinct GPUs with ROCm. The release includes model weights from each training stage, data mixtures, code, and a 16B-total, 2.8B-active architecture meant to show that open AMD-based training stacks can compete. Read more

Google added natural voice control to the Gemini app for macOS. Long-pressing the Fn key lets users dictate polished text into any app, and optional Gemini reasoning can summarize files, rewrite selected text, or edit images based on what is on screen. Read more

Hugging Face published a technical timeline of a July 2026 agent intrusion that produced about 17,600 actions over several days. The write-up explains how the agent chained ordinary security weaknesses across trust boundaries, and why defenders need tighter isolation, short-lived credentials, and better cross-system detection. Read more

Unlock next-level video creation with MiniMax H3, the powerful new model that accepts text, image, audio, and video as references through its Omni Reference system. Generate, edit, and transform content with unprecedented flexibility. Deliver native 2K video up to 15 seconds, purpose-built for film, advertising, business, and more. Try it today!

Vivix-A1 is a real-time model for full-body AI characters that can listen, speak, move, and act inside open worlds. It supports voice, text, and image inputs during a live interaction, so a character can be redirected while it is already talking or moving. Read more

ByteDance Seed launched Seedance 2.5, a new audio-video generation model built for longer 30-second stories. It gives creators tighter reference control, smoother motion, and editing tools like green-screen and white-model control for more production-style workflows. Read more

BeingBeyond introduced Being-H0.8, a tactile-aware embodied foundation model for robots. It learns from more than 500,000 hours of egocentric human video, adds pseudo-touch labels through TactoHand, and uses a shared hand/action representation for human hands, robot hands, and grippers. Read more

Wonder is a video world model that turns an image or video into an interactive world users can navigate. It keeps memory of previous views, reveals unseen areas, and supports up to one-minute rollouts at constant latency with 16 FPS generation. Read more

Nyra Labs released CrisperWhisper 2.0 for verbatim speech recognition. The model can switch between word-for-word transcripts and cleaned-up intended transcripts, supports multiple languages, and reports precise word-level timing for speech that includes fillers, stutters, and vocal sounds. Read more

PhiZero proposes a world model built around a learned physical language instead of plain natural language. It predicts future scene changes as compact transition tokens first, then renders them into video, which could help models reason about physics before making pixels. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like “um,” fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

ID-V2V is a Netflix and Eyeline Labs research project for restyling videos while preserving a person’s identity and performance. It uses relit face regions, facial normals, depth, and edited keyframes so style changes can spread through a clip without losing expressions, gaze, or lip sync. Read more

PRISM is a robot-control method that learns useful physical cues from signals a robot already has. By modeling polynomial interactions in proprioception, it improves humanoid survival, manipulation success, and long-horizon LIBERO tasks without adding new sensors. Read more

ReDesign turns a flat design screenshot back into editable design structure. The ECCV 2026 project decomposes an image into text, shapes, images, groups, and layer order, then exports an editable JSON hierarchy for recovering lost design files. Read more

ShadowDancer teaches video world models to replay actions from demonstrations instead of labels or text commands. It learns from paired videos that share the same dynamics but change the scene, character, or lighting, then transfers those actions into new worlds and longer rollouts. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

Google introduced Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber for building faster, cheaper AI agents at scale. 3.6 Flash cuts output token use versus 3.5 Flash while improving coding and multimodal work, Flash-Lite targets speed, and Flash Cyber pairs with CodeMender for security tasks. Read more

OpenAI is rolling out Health in ChatGPT for U.S. users 18 and older on web and iOS. Users can connect Apple Health and supported medical records so ChatGPT can compare results, summarize changes, and answer health questions with personal context, while OpenAI says connected health data is not used to train foundation models or target ads. Read more

OpenAI and Hugging Face disclosed a security incident triggered during an internal cyber-capability evaluation. OpenAI says evaluation models found a way out of the isolated test setup, reached Hugging Face systems while pursuing benchmark answers, and both companies are now investigating and tightening controls. Read more

NVIDIA and other large tech companies published a letter urging U.S. policymakers to avoid premature restrictions on open-weight AI models. They argue that open models help startups, researchers, and defenders inspect systems, compete, and build AI without lock-in. Anthropic is the only large AI lab that has not signed this letter. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

Black Forest Labs previewed FLUX 3, a coming multimodal model for image, video, audio, and action prediction. Details are still limited, but the pitch is a single creative model that can make realistic outputs across styles instead of separate tools for each media type. Read more

Alibaba launched Qwen-Image-3.0, its third-generation image generation model focused on useful, information-dense visuals. The official page was thin in direct fetches, but Qwen’s release materials describe stronger small-text rendering, richer layouts, realistic detail, and knowledge-heavy images such as documents and diagrams. Read more

Poolside released Laguna S 2.1, a 118B-parameter coding model with only 8B active parameters per token. It supports up to a 1M-token context and is designed for long-horizon software work, with Poolside highlighting strong coding benchmarks and full evaluation trajectories for every final trial. Read more

Microsoft’s Mage team introduced Mage-Flow, a compact 4B image generation and editing model family built for fast high-resolution work. Its Turbo versions generate 1024-square images in about 0.59 seconds and edits in about 1.02 seconds on a single A100, showing how smaller image models can still feel interactive. Read more

GLM is Z.AI’s flagship open-source model built for long, difficult engineering tasks like coding, debugging, and autonomous agent work. It delivers top-tier performance while being dramatically cheaper and faster than frontier closed models. Get 10% OFF HERE.

Nanbeige4.2-3B is a compact open-weight agent model designed for tool use, coding, reasoning, and local assistant workflows. Its Looped Transformer reuses layers to increase capacity without adding parameters, and the model card says it beats larger Qwen and Gemma models on several agent and code benchmarks. Read more

OpenAI’s docs now describe ChatGPT Voice for Chat, Work, and Codex in the ChatGPT desktop app. The feature lets users talk through ideas, start or steer tasks, check progress, and use voice dictation when they only need speech turned into prompt text. Read more

Google Quantum AI showed a reinforcement-learning approach that lets a quantum computer adjust while it runs. The system learns from error-correction signals to keep thousands of control settings stable, which matters because useful quantum programs may need to run for days or months without stopping. Read more

Stanford Medicine researchers used AI to find BRP, a natural peptide that may reduce appetite and body weight like semaglutide. In animal studies, BRP cut food intake and fat while avoiding clear signs of nausea, constipation, and major muscle loss, but it still needs human testing. Read more

NVIDIA Research introduced SANA-Video 2.0, an efficient video generation model built to make high-quality video on a single GPU. The project reports 720p five-second generation in about 13 seconds on one H100 for its optimized 5B pipeline, which could make video models cheaper to run and test. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like “um,” fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

ShotPlan is a research system for making multi-shot videos with precise scene changes and camera moves. It adds learnable planning tokens to a text-to-video diffusion model so creators can ask for hard cuts, cross-fades, and timed camera motion instead of hoping the model guesses the edit. Read more

Stanford-led researchers introduced masked visual actions as a way to control video world models across different robots and objects. Instead of feeding a robot-specific action format, the model uses a partial action video and predicts what happens next, which could help plan or evaluate robot behavior before real-world execution. Read more

Open Dreamer is an open effort to reproduce frontier-style world model training and share the lessons along the way. The team is open-sourcing the model and training code, starting with CoinRun as a smaller testbed before aiming at more complex Minecraft-like world modeling. Read more

HOMIE is a video personalization framework focused on keeping both people and objects consistent in generated videos. It targets cases where humans interact with objects, logos, or multi-view references, aiming to preserve identity, text, and object relationships better than earlier subject-driven video methods. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

Thinking Machines released Inkling, an open-weights multimodal model with 975B total parameters and 41B active parameters. The model supports up to 1M tokens and is built for customization through Tinker, including fine-tuning workflows that developers can run directly. Read more

OpenAI introduced GPT-Red, an internal automated red-teaming model trained to find prompt-injection failures. OpenAI says attacks generated by GPT-Red helped make GPT-5.6 much more robust while preserving normal capabilities. Read more

NVIDIA released Nemotron 3 Embed, a family of open embedding models for RAG, agent memory, code retrieval, and search. The 1B and 8B models are designed to improve retrieval quality while giving teams practical options for production deployment. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

OpenAI and Work Louder introduced Codex Micro, a $230 hardware controller for agentic work. Its keys, dial, joystick, and RGB status lights are meant to keep Codex actions and agent state visible while developers work. Read more

PrismML released Bonsai 27B, a compressed multimodal model based on Qwen3.6 27B that it says can run on a phone. The 1-bit version is 3.9GB and the ternary version is 5.9GB, which matters for private, local agents and offline use. Read more

Qualcomm AI Research introduced MobileWan, a 5B video diffusion system designed to run on commercial mobile hardware. It generates 5-second 480x832 videos at 16 FPS in about 20 seconds, narrowing the gap between mobile and server video models. Read more

Wan Streamer v0.3 turns real-time video generation into a 'world + event stream' problem. That lets the model keep a scene coherent while speech, motion, camera movement, and actions change at around 200 ms model-side latency. Read more

GLM is Z.AI's flagship open-source model built for long, difficult engineering tasks like coding, debugging, and autonomous agent work. It delivers top-tier performance while being dramatically cheaper and faster than frontier closed models. Get 10% OFF HERE.

Google DeepMind researchers showed that video generation backbones can be repurposed as general-purpose vision learners. GenCeption uses a text-steered feed-forward model for tasks like depth, segmentation, camera pose, and 4D keypoints with strong data efficiency. Read more

Schema is a new harness that reports near-ceiling scores on the ARC-AGI-3 public set without changing model weights. The authors say the results are public-set and self-reported, but the core idea is important: forcing models to build and test executable world models can matter as much as the model itself. Read more

Google released GNM, an open ecosystem for high-fidelity parametric human models, starting with GNM Head. It gives researchers and builders a permissively licensed 3D head model with controllable identity, expression, pose, eyeballs, teeth, and tongue. Read more

NVIDIA's ARDY generates controllable 3D human motion in real time from text and movement constraints. It is built for interactive animation, games, simulation, and humanoid robot control, where offline motion tools are too slow. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like "um," fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

Alibaba's Wan-Dancer generates minute-scale 720p dance videos that stay synchronized with music. The project uses global keyframe planning and local refinement to reduce the drift and identity changes that usually appear in long AI dance clips. Read more

Motion4Motion transfers movement between very different subjects without a shared skeleton or new training. It tracks dense motion flow from a source video and injects it into a frozen video diffusion model, enabling transfers like human-to-animal or a walking table demo. Read more

Mirelo and Kyutai open-sourced MuScriptor, an audio-to-MIDI model for full songs. It can separate a mix into MIDI tracks for voice, drums, bass, keys, and other instruments, which makes dense music easier to study or edit in a DAW. Read more

Lucida is an MIT-licensed background-removal model tuned for cases where common matting tools fail. It focuses on transparent objects, camouflage, text, glow effects, and illustrations, with weights on Hugging Face and a browser demo. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

OpenAI is rolling out GPT‑Live, a full-duplex voice model that can listen and speak at the same time for more natural conversation. It keeps talking while delegating search and hard reasoning to frontier models, and is rolling out globally in ChatGPT before an API release. Read more

OpenAI released GPT-5.6, a new family of smarter and more efficient AI models including the flagship Sol. It delivers top performance in coding, science, knowledge work, and complex tasks while using fewer tokens and costing less than before — with special “ultra” mode that teams up multiple agents for really hard jobs. Smaller models like Terra and Luna make advanced AI more affordable for everyday use. Read more

Meta’s Muse Spark 1.1 is a multimodal reasoning model built for agents, computer use, coding, and long-running workflows. Its 1-million-token context and parallel subagent orchestration are now available through a public-preview Meta Model API. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

xAI’s Grok 4.5 is a model aimed at coding, agentic tasks, and knowledge work, with a focus on engineering performance and efficiency. It is available in Grok Build, Cursor, and the xAI console, where xAI says it serves at 80 tokens per second and costs $2 per million input tokens and $6 per million output tokens. Read more

Meta has launched Muse Image and previewed Muse Video as its first media-generation models from Meta Superintelligence Labs. Muse Image can search, write code, self-refine, edit and combine references, while Muse Video targets high-fidelity video with native audio. Read more

ByteDance’s Seedream 5.0 Pro is a multimodal image generator built for advanced reasoning and professional visual production. It focuses on text-rich infographics, sketch- and annotation-guided editing with layer separation, photorealistic lighting and skin, and native support for a dozen languages. Read more

ABot-World turns a single RTX 5090 desktop GPU into an interactive world simulator that responds to user actions. The project reports 720p generation at 16 fps with 1.2-second latency and 19 GB of memory, enabling open-ended rollouts instead of fixed-length videos. Read more

GLM is Z.AI’s flagship open-source model built for long, difficult engineering tasks like coding, debugging, and autonomous agent work. It delivers top-tier performance while being dramatically cheaper and faster than frontier closed models. Get 10% OFF HERE.

AlayaWorld is a 15-billion-parameter world model that generates playable video worlds at 720p and 24 fps for more than a minute. It combines a 3D scene cache with frame-history memory so users can move the camera, switch prompts mid-scene, and revisit places without the world changing as much. Read more

MIRA is a 5-billion-parameter model that simulates four-player Rocket League-style matches directly from video and player actions. Built by General Intuition and Kyutai with Epic Games, it runs a shared 2v2 world at 20 fps without a traditional game or physics engine, and the team is releasing code and data. Read more

Wan Streamer v0.2 raises real-time generated video from 192×336 to 640×368 while keeping about 200 ms of model-side latency. Its thinker–performer setup uses one GPU for perception and state updates plus parallel GPUs for the higher-resolution video, making scene-grounded mid-shot agents possible at 25 fps. Read more

Tencent’s open-source Hy3 is a smaller agent model tuned for reasoning, coding, office work, long contexts, and reliable tool use. Tencent reports a 2.67/4 score from a 270-expert blind evaluation, lower hallucination and multi-turn error rates than its preview, and Apache 2.0 weights with API pricing starting at 1 RMB per million input tokens. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like “um,” fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

ProxyPose turns one clicked pixel in a monocular video into a full 6-DoF motion trajectory by first generating a proxy video. That simple interface lets it track rigid objects, cameras, faces, event-camera footage, and single-photon video, though fast motion and reflective or textureless surfaces can still cause drift. Read more

PixWorld unifies 3D scene reconstruction and generation in one pixel-space diffusion model. Instead of hiding the scene inside a latent code, it trains directly on rendered views and adds geometry-aware supervision, aiming for more faithful 3D structure with fewer representation-loss problems. Read more

SeFi-Image is a text-to-image foundation model that denoises semantic structure slightly ahead of texture details. Its 1B, 2B, and 5B versions use that two-stream approach to improve layout, long text, and visual fidelity while the 5B model was trained with 125,000 A800 GPU hours. Read more

A study evaluates a humanoid robot for laparoscopic surgery using teleoperation, benchtop tests, user studies, and live porcine experiments. The work shows that general-purpose humanoid hardware can perform surgical tasks, but it also highlights the precision, control, and safety gaps that must be closed before clinical use. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

Anthropic launched Claude Sonnet 5 as its most performant Sonnet model yet. It is now the default model for Free and Pro plans, and is available in Claude Code and the API with lower-cost pricing aimed at Opus-like tool use, coding, and knowledge work. Read more

Google released Nano Banana 2 Lite and rolled Gemini Omni Flash to developers. Nano Banana 2 Lite targets four-second, low-cost image generation, while Omni Flash adds multimodal video generation and conversational editing in Google AI Studio and the Gemini API. Read more

Anthropic says Claude Fable 5 and Mythos 5 access has been restored after export controls were lifted. Fable 5 returns globally with a new safety classifier for a reported safeguard bypass, while Mythos 5 access remains tied to approved defensive-cyber partners. Read more

Vidu launched Vidu S1 for real-time interactive AI video characters. Users can create a character from an image, choose or clone a voice, and run live voice-driven calls with endless 540p video generation at 25fps or higher. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

InternScience released Agents-A1, a 35B MoE model for long-horizon agent work. It supports a served 256K context and is evaluated across search, engineering, scientific research, instruction following, and tool calling. Read more

ComfyUI introduced Comfy MCP in public beta so agents can operate Comfy workflows. The connector gives Claude, Codex, Cursor, and other agents access to image, video, 3D, and audio models plus ComfyUI workflow search, execution, and reuse. Read more

Meta released Brain2Qwerty v2 research for decoding words from non-invasive brain recordings. The system uses MEG brain signals and language-model context to reach 61% word accuracy overall and 78% for the best participant, while releasing training code and a dataset. Read more

NVIDIA GEAR introduced ASPIRE, a self-improving robotics system that writes and repairs code-as-policy programs. It inspects rollout traces, fixes failures, saves reusable robot skills, and improves tasks like two-arm handover from 20% to 92% in Robosuite. Read more

GLM is Z.AI’s flagship open-source model built for long, difficult engineering tasks like coding, debugging, and autonomous agent work. It delivers top-tier performance while being dramatically cheaper and faster than frontier closed models. Get 10% OFF HERE.

NVIDIA Isaac researchers introduced CHORD for learning dexterous robot manipulation from human demonstrations. Instead of only matching contact positions, CHORD matches the forces and torques that move objects, reaching 82.12% average success across 1,831 benchmark tasks and transferring to real robots. Read more

Perceptive BFM adapts human motion references to robot terrain in real time. A single policy follows raw motions like flips or fall recovery while using perception to choose footholds, clearance, posture, and contact timing for the terrain it sees. Read more

OmniContact turns contact-flow meta-skills into general humanoid loco-manipulation behaviors. Its dataset and live MuJoCo viewer show a humanoid carrying, pushing, kicking, recovering, and chaining tasks with closed-loop monitoring and VLM instruction support. Read more

LiveEdit brings diffusion-based video editing closer to real-time streaming. The ECCV 2026 project distills a bidirectional editing model into a causal streaming editor, using cache reuse to preserve backgrounds while reaching 12.66 FPS. Read more

Representation Distribution Matching trains a one-step image generator without an online teacher or adversary. Its iRDM approach matches generated and real feature distributions across frozen encoders, beating its four-step FLUX.2 teacher on GenEval and PickScore after about 90 H200 GPU-hours. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like “um,” fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

MrFlow speeds up text-to-image diffusion without training a new model. It samples cheaply at low resolution, upscales, re-encodes, and refines at high resolution, reporting more than 10x end-to-end speedup on Qwen-Image while preserving quality. Read more

PhysiFormer is a diffusion transformer for simulating 3D object mechanics in world coordinates. It predicts full mesh trajectories for rigid and elastic objects from initial positions, velocities, and materials, offering a learned alternative that can be faster than physics simulators after training. Read more

LUNA animates 3D human avatars without relying on traditional linear blend skinning. The ECCV 2026 project maps images, keypoints, sketches, and unseen characters into 3D Gaussian deformations, aiming for realistic zero-shot animation from 2D controls. Read more

ViDiHand uses video diffusion features to reconstruct smooth 4D hand motion. By fine-tuning a hand-aware branch while freezing the base diffusion transformer, it improves occlusion robustness, accuracy, and temporal smoothness across egocentric benchmarks. Read more

MuSViT is a foundation vision model built specifically for sheet music. It was pretrained on 9.7 million IMSLP pages and outperforms general-purpose encoders on music recognition, symbol detection, and score difficulty tasks. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!

Note that most of the examples are visual. See my Youtube video for the best experience.

OpenAI and Broadcom are working on Jalapeño, a custom inference chip built to run AI models more efficiently at huge scale. Instead of relying only on general-purpose chips, OpenAI wants hardware tuned for serving AI responses faster and cheaper. Read more

ByteDance released Seed2.1, a new family of AI productivity models built for agents, coding, tool use, and multimodal work. The Pro and Turbo versions are designed to handle real office tasks, long coding jobs, debugging, file processing, and complex visual or video understanding. Read more

IBM says it has built the world’s first sub-1 nanometer chip technology, using a new 3D “nanostack” transistor design. If it scales, this could help keep chips getting smaller, faster, and more energy-efficient even as old chip designs hit physical limits. Read more

Do you prefer to watch instead of read? Check out this video covering the top AI news this week:

Meta introduced Autodata, a system where AI agents act like data scientists to create better training and evaluation data. That matters because better data can make models smarter, and Meta shows the agent can improve its own data-making recipe over time. Read more

Sakana Fugu is a multi-agent system that works like one model by dynamically choosing and coordinating top AI models for hard tasks. Instead of betting on one provider, it routes work across different models so users can get stronger results from a single API. Read more

Wan Streamer is a real-time AI model that can listen, see, think, talk, and respond with video at about 25 frames per second. It puts audio, video, and text into one streaming Transformer, which could make future AI video agents feel much more live and natural. Read more

The U.S. reportedly allowed Anthropic to release a model called Mythos to trusted American organizations. This is important because advanced AI access is becoming tied to national security rules, not just normal product launches. Read more

GLM is Z.AI’s flagship open-source model built for long, difficult engineering tasks like coding, debugging, and autonomous agent work. It delivers top-tier performance while being dramatically cheaper and faster than frontier closed models. Get 10% OFF HERE.

HIW-500 is a 500-hour humanoid robot dataset collected in real homes, not clean lab setups. It gives robots examples of messy household tasks with different layouts, lighting, clutter, and human control styles, which is exactly what robots need to learn real-world behavior. Read more

Aleph is building ultrasound-based brain imaging that aims to get MRI-like detail without drilling into the skull or using a giant scanner. The idea is to read blood-flow changes caused by brain activity, which could make future brain-computer interfaces less invasive and more practical. Read more

DomainShuttle is a text-to-video system that keeps the same subject consistent while moving it across very different worlds and styles. That means a real person, object, or character can be turned into fantasy, storybook, or realistic video scenes without losing their identity. Read more

DanceOPD is a training method that distills multiple image-generation abilities into one student model more cleanly. It learns by asking the right expert “field” during the student model’s own rollout, which can make smaller models inherit editing, style, and text-to-image skills more reliably. Read more

PerceptionDLM is a multimodal diffusion language model that can describe multiple image regions in parallel. Instead of captioning one region after another, it uses diffusion-style parallel decoding, which can make detailed visual understanding faster. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like “um,” fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

Arbor gives 3D asset generation direct shape controls using constraint meshes. Artists can mark where geometry should exist, where it should stay empty, and what surfaces should touch, making AI-generated 3D objects easier to control. Read more

Lift4D reconstructs moving 3D objects from a single normal video, even filling in parts the camera never directly sees. It combines single-view 3D priors with test-time optimization, which helps with hard real-world scenes that have occlusion and big motion. Read more

UN-0 is an experimental image generator based on coupled oscillators instead of the usual neural-network image pipeline. It is a more unusual computing approach, but the point is exciting: new hardware-inspired methods might eventually generate images in cheaper or stranger ways. Read more

Ornith 1.0 is an AI release from DeepReinforce focused on reinforcement-learning style progress for smarter agents. The public page is sparse, but the project points toward systems that learn through feedback rather than only copying examples. Read more

Meet Fish Audio, the most expressive AI voice model. Core features include text-to-speech, instant voice cloning, and a library of over 2 million voices. Create natural, emotionally rich AI voices built for your workflows. Try it for free.

Also check out our Youtube for more AI news & reviews!