Quanta AI Labs

28 / Jul 6 – Jul 12, 2026 · 35 RELEASES

AI News Week 28, 2026

Frontier Flood + World-Model Wave. Frontier releases land in a cluster: xAI Grok 4.5 (Opus-class, 500K context, ~2x token efficiency at $2/$6 pricing), OpenAI GPT-Live full-duplex voice models, Meta Superintelligence Labs Muse Spark 1.1 (1M-context agentic reasoning) plus Muse Image and Muse Video, Tencent open-sources Hunyuan Hy3 (295B MoE, 21B active, Apache 2.0), and an unverified leak surfaces a "Claude Fable 5"/"Mythos 5" family. Image generation heats up with ByteDance Seedream 5.0 Pro, Reve 2.1 (native 4K), and SeFi-Image (semantic-first diffusion). Real-time world models surge: ABot-World (720p/16fps on one RTX 5090), AlayaWorld (minute-long controllable video), Wan Streamer v0.2 (duplex audio-video), General Intuition MIRA (playable 2v2 Rocket League sim + Rocket Science dataset), and Ant Group LingBot-World 2.0 (hour-long 720p/60fps). Huge agent-tooling wave: Alook and Omnigent (multi-agent workforces/meta-harness), Mission Control fleet dashboard, BMad Method, Squid, OmniAgent, Memento (memory-based continual learning, 79.4% GAIA), Alibaba Page Agent, OpenScience, plus LangChain OpenWiki doc agent, OfficeCLI, Claude Video, last30days, and OpenAI's Codex plugin for Claude Code. AI filmmaking previs: Blockout and Motion Previs Studio. Security: Strix autonomous pentester. Research: UCSD Humanoid Surgeon (in-vivo robot laparoscopy), ProxyPose, PixWorld.

Language Models (3)

  • Grok 4.5

    xAI · Text Generation

    xAI Opus-class model built for coding and agentic knowledge work, landing 4th on the Artificial Analysis Intelligence Index with a 500K-token context, tool use and vision, and roughly 2x token efficiency. Priced at $2/M input and $6/M output — over 60% cheaper than Opus 4.8 or GPT-5.5 — scoring 62% on DeepSWE 1.0 and 33% on tau3-Banking.

  • Muse Spark 1.1

    Meta · Reasoning

    Multimodal reasoning model from Meta Superintelligence Labs built for agentic tasks, with a 1-million-token context window and gains in tool use, computer interaction, coding, and multimodal understanding. Available via the new Meta Model API in public preview and as "Thinking" mode in the Meta AI app.

  • Hunyuan Hy3

    Tencent · Text Generation

    Tencent flagship open-source Hunyuan 3.0 model, a Mixture-of-Experts LLM with 295B total parameters (21B activated per inference) and a 256K-token context window, released under Apache 2.0. Priced at ~$0.15/M input and ~$0.59/M output, with agent capabilities integrated across WeChat and Tencent products.

Vision & Image (4)

  • Riddle

    Maxime Rivest · OCR & Document

    Interactive AI diary app for the reMarkable Paper Pro e-ink tablet. Users write by hand; the text fades and the device replies with an AI-generated response rendered in flowing handwriting via handwriting recognition.

  • Muse Image & Muse Video

    Meta · Text-to-Image

    Meta's most advanced image generation model (Muse Image, available now) is an agentic system with tool use (web search, code execution) and self-refinement for faithful instruction-following, precise editing, and multi-reference composition. Muse Video (coming soon), built on the same foundation, adds integrated audio and ranks third on Arena for text-to-video.

  • Reve 2.1

    Reve · Text-to-Image

    Image generation model billed as the top 4K model in the world, excelling at dense scenes, fine text, and intricate structure rendered at native 4K. Adds enhanced multilingual text rendering and precision editing across individual elements.

  • SeFi-Image

    SeFi-Team · Text-to-Image

    Text-to-image foundation model using "semantic-first" diffusion, where semantic and texture latents follow separate denoising trajectories (semantic ahead of texture) to improve fidelity; released in 1B, 2B, and 5B variants with competitive results on GenEval, LongTextBench, and visual text generation.

Video & Animation (7)

  • Claude Video

    bradautomates · Video Understanding

    Claude Code skill/plugin that lets Claude analyze videos by downloading them, extracting frames, and transcribing audio. Accepts YouTube/TikTok/Vimeo URLs or local files with tunable frame-extraction modes, pulls native captions or falls back to Whisper; built on yt-dlp and ffmpeg, MIT-licensed.

  • Blockout

    Wasserman Productions · Video Editing

    Open-source desktop previs app for AI-native filmmaking that lets users stage grey-box scenes and choreograph camera moves and character blocking, then export motion-reference packages for AI video generators like Seedance, Veo, Kling, LTX, and Wan.

  • Motion Previs Studio

    Sam Wasserman · Video Editing

    Open-source desktop app that extracts pose, depth, camera movement, and control layers from video footage for AI-film previsualization workflows compatible with ComfyUI, Blender, and Runway.

  • ABot-World

    Alibaba AMAP CV Lab · Real-time Video

    Action-conditioned interactive world model that generates real-time, open-ended video rollouts steered by user actions, producing 720p output at 16 FPS on a single consumer NVIDIA RTX 5090 GPU.

  • AlayaWorld

    ALAYA Lab · Real-time Video

    Interactive autoregressive world model generating minute-long video with user control over camera movement and narrative events, combining a 3D spatial cache with temporal memory for consistency, running at 720p and 24 FPS with real-time prompt switching and camera modulation.

  • MIRA

    General Intuition · Real-time Video

    5B-parameter latent-diffusion multiplayer world model that simulates full 2v2 Rocket League matches in real time at 20 FPS on a single GPU from four players' actions, with no physics or rendering engine; paired with a 600M video codec and released alongside the "Rocket Science" dataset (~1,000 hours of matches) plus open training/inference code.

  • AI Video Analysis Tool

    Vinay Bajjuri · Video Understanding

    Personal demo tool that analyzes uploaded videos with generative AI to produce structured insights (participant identification, position/event timelines, dominance-phase estimation, performance summaries), shown on a Brazilian Jiu-Jitsu clip; spans OpenCV, vision LLMs, transcription, semantic indexing, and agent orchestration.

Audio & Speech (1)

  • Wan Streamer v0.2

    Tongyi Lab (Alibaba) · Voice Agents

    End-to-end Transformer for real-time duplex audio-visual conversation that perceives and responds with synchronized speech and video as a single causal stream; v0.2 raises output resolution from 192x336 to 640x368 at 25 FPS while holding ~200ms model-side (~550ms total) latency via a Thinker-Performer architecture, and can embody any natural-language-described character.

3D & Spatial (2)

  • ProxyPose

    University of Toronto & Vector Institute · Depth Estimation

    Method for recovering full 6-DoF pose trajectories from monocular video via video-to-video translation, rendering a colored proxy cube that mirrors a query point's rigid motion and then applying contour fitting and PnP optimization; uses a LoRA-finetuned video generative model for temporal consistency.

  • PixWorld

    Nanyang Technological University & AISphere · 3D Generation

    Single end-to-end pixel-space diffusion model that unifies 3D scene generation and reconstruction, processing multi-view inputs through a two-stream diffusion transformer to output pixel-aligned 3D Gaussians with a geometry-perception loss, reaching ~0.6s scene generation after distillation.

Agents & Automation (14)

  • Memento

    Memento-Teams · Agent Framework

    Memory-based continual-learning framework that lets LLM agents improve from experience without updating model weights, using case-based reasoning over stored success/failure trajectories and a planner-executor architecture with MCP tools. Reports 79.40% on the GAIA test set.

  • Alook

    alookai · Agent Orchestration

    Open-source, self-hosted platform for running teams of collaborating AI coding agents as an autonomous workforce, giving each agent its own email address, role, Kanban board, and scheduling. Supports Claude Code, Codex, and OpenCode; built with TypeScript/Next.js/Cloudflare Workers, Apache-2.0.

  • Squid

    iusztinpaul · Coding Agents

    Claude Code plugin that automates software development through a 5-agent pipeline (Product Architect, Software Engineer, Tester, PR Reviewer, On-Call Engineer) with two human approval gates, turning feature specs into reviewed pull requests. Includes scaffolding and standalone skills for Python, TypeScript, and Go stacks.

  • BMad Method

    BMad Code · Agent Framework

    Open-source AI-driven agile development framework with 12+ specialized agent personas (PM, architect, developer, UX, etc.) and 34+ workflows spanning ideation through deployment. Adjusts planning depth to project scope; MIT-licensed, npm-installed, works with Claude Code, Cursor, and other AI IDEs.

  • Omnigent

    omnigent-ai · Agent Orchestration

    Open-source agent meta-harness providing a unified orchestration layer over multiple AI coding agents (Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, custom), with cross-device session sync, cloud sandboxing via Modal/E2B/Kubernetes, and policy governance for approvals and spend caps. Agents are authorable in YAML.

  • Strix

    usestrix · Agent Framework

    Open-source AI penetration-testing tool that autonomously finds and validates application vulnerabilities via dynamic testing and proof-of-concept exploits. Uses a multi-agent architecture to cover OWASP Top 10 and beyond, with auto-fix and compliance reporting.

  • Page Agent

    Alibaba · Web Agents

    Open-source JavaScript in-page GUI agent that controls web interfaces through natural-language commands. Runs directly inside the webpage using text-based DOM manipulation, requiring no browser extension.

  • OpenScience

    Synthetic Sciences · Agent Framework

    AI workbench that automates the scientific research process end-to-end, from literature review and hypothesis formation through code execution, experimentation, and write-up. Runs in a browser-based workspace integrating multiple AI models and scientific databases.

  • Mission Control

    Builderz Labs · Agent Orchestration

    Open-source, self-hosted dashboard for orchestrating fleets of AI agents. Lets teams dispatch tasks, run multi-agent workflows, monitor spend, and govern operations from one interface, backed by SQLite with zero external dependencies.

  • OmniAgent

    CubeStar1 · Agent Framework

    Autonomous multi-agent AI system combining specialized agents with automation tools for complex task execution, including document processing, web automation, code execution, and query retrieval. Built on LLM agents and MCP servers; won Bajaj Finserv HackRX 6.0.

  • last30days

    mvanhorn · Coding Agents

    Agent skill that researches a topic across Reddit, X, YouTube, Hacker News, Polymarket, and the web, then synthesizes a grounded summary ranked by real engagement metrics. Works with Claude Code, Cursor, Copilot, and 50+ other agent hosts.

  • Intelligent Trading Bot

    asavinov · Agent Framework

    Python automated trading system that uses machine learning and feature engineering to generate crypto trading signals, separating offline model training from online signal generation. Supports backtesting, technical indicators, and Binance/Telegram integration.

  • AI Ad Agent Playbook

    Creatify · Agent Framework

    Notion resource/playbook outlining how to build and use an AI ad agent for automated advertising creative — a guide rather than a model or product launch.

  • Humanoid Surgeon

    UC San Diego ARCLAB · Robotics

    In-vivo feasibility study evaluating whether general-purpose humanoid robots can perform teleoperated minimally invasive laparoscopic surgery, validated across benchtop tests, dry-lab surgeon studies, and a live porcine cholecystectomy, benchmarked against platforms like the da Vinci system.

Developer Tools (3)

  • OpenWiki

    LangChain · AI Infrastructure

    Open-source CLI from LangChain that auto-generates and maintains codebase documentation for AI agents, with a code mode for repo docs and a personal mode that builds a local wiki from Gmail, Notion, X, web search, and Hacker News. Generates AGENTS.md/CLAUDE.md and integrates with GitHub Actions, GitLab CI, and Bitbucket Pipelines.

  • OfficeCLI

    iOfficeAI · AI Infrastructure

    Open-source CLI tool that lets AI agents create and edit Word, Excel, and PowerPoint documents without Microsoft Office installed. Includes an HTML rendering engine for document visualization, 350+ Excel functions, and native pivot table generation.

  • Codex Plugin for Claude Code

    OpenAI · Prompt Engineering

    Official OpenAI plugin that lets Claude Code users delegate code reviews and tasks to Codex from within their existing workflow via slash commands like /codex:review and /codex:rescue.

Research & Safety (1)

  • Claude Fable 5 (System Prompt Leak)

    asgeirtj/system_prompts_leaks · Alignment & Safety

    Leaked system prompt referencing "Claude Fable 5," described as the first model in a new Claude 5 family and part of a "Mythos-class" tier positioned above Claude Opus. The leak indicates Fable 5 is the safety-restricted public model while a companion "Claude Mythos 5" offers the same capabilities without restrictions to approved organizations. Unverified.

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.