Quanta AI Labs

25 / Jun 15 – Jun 21, 2026 · 26 RELEASES

AI News Week 25, 2026

Open Audio-Video + Science Foundation Models. Lightricks ships LTX-2 — first open-weights (Apache 2.0) DiT foundation model jointly generating synchronized video + audio (dialogue, music, SFX) in one pass at native 4K/50fps/20s (19B params: 14B video + 5B audio) — plus the LTX-2 trainer for LoRA/full fine-tuning. Boogu-Image-0.1 (10B Apache-2.0, ultra-dense bilingual text rendering). OpenAI pairs GPT-5.4 with Molecule.one's agent to improve a real chemistry reaction across 10,080 wet-lab runs (mean yield 16.6%→25.2%) + ships Codex Record & Replay (demonstrate a workflow once, Codex learns a reusable skill). LOGOS — first multi-domain generative foundation model for natural sciences (proteins, molecules, materials, reactions in one autoregressive transformer). Midjourney's surprise hardware pivot: Midjourney Medical 'Scanner' ultrasonic-CT full-body imaging. World models: AMAP DreamX World, PermaVid (disentangled RGB+depth memory), OmniDirector camera-motion cloning. Image/style: StoryGen visual storytelling, TeleStyle V2 style transfer, StoryGen-Atelier (Gemini+Veo). Agent infra: Agent Reach (multi-platform agent access), OpenMontage (agentic video production), Agentic DB (Postgres agent memory), Karpathy's LLM Wiki pattern, Google Open Knowledge Format, PixelRAG (screenshot-based RAG). Evals: DeepEval, iFixAi (agent misalignment audit). Robotics: UME-Exo teleoperation exoskeleton, Moya humanoid. Tools: Headroom (60-95% token compression), Unsloth GLM-5.2 guide, palmier-pro macOS AI editor, PixlStash, optimizerDuck, AI-ML Companion, Gorest spritesheet generator.

Vision & Image (2)

  • StoryGen

    haoningwu3639 · Text-to-Image

    Visual storytelling system generating coherent image sequences from text narratives via a two-stage latent diffusion pipeline: pretrain Stable Diffusion self-attention for style, then train a Visual-Language Context Module to condition on prior story frames. Includes a data pipeline with keyframe extraction, dedup, segmentation, inpainting and DTW subtitle alignment.

  • TeleStyle V2

    witcherofresearch · Image Editing

    Content-preserving style transfer model that generates stylized outputs with content and style consistency across four reference combinations (realistic/stylized pairings), using self-distillation data synthesis and distribution matching distillation. Retains general image editing and text-guided editing alongside style transfer.

Video & Animation (6)

  • StoryGen-Atelier

    0xsline · Text-to-Video

    Text-to-video pipeline that generates storyboard scripts + images via Gemini, produces transition clips between keyframes with Vertex AI Veo, and assembles the final cut with FFmpeg. React/Vite/Mantine frontend, Node.js/Express + better-sqlite3 backend.

  • LTX-2

    Lightricks · Text-to-Video

    First open-weights (Apache 2.0) DiT audio-video foundation model that jointly generates synchronized video and audio (dialogue, music, SFX) in a single pass at up to native 4K/50fps and 20s duration, using an asymmetric dual-stream DiT with bidirectional cross-attention. 19B params (14B video + 5B audio).

  • Palmier Pro

    palmier-io · Video Editing

    Swift-native macOS video editor with built-in generative AI (Seedance, Kling, Nano Banana Pro) and MCP server integration, letting users and AI agents collaboratively generate and edit video on a timeline. Open source (GPLv3), targeting macOS 26 Tahoe on Apple Silicon.

  • DreamX World

    AMAP / Alibaba · Text-to-Video

    General-purpose world model that generates interactive, controllable environments from action and event prompts, combining video generation with spatial memory, long-horizon rollouts and reinforcement learning. Targets explorable, transformable high-fidelity worlds across realistic and imaginative settings.

  • PermaVid

    ys-imtech · Text-to-Video

    Video generation framework that maintains a disentangled multi-modal context memory — an RGB bank for semantic appearance and a depth bank for geometric structure — to keep videos consistent across camera viewpoints and edits, using memory-guided generation with mixed-modality feature fusion.

  • OmniDirector

    ymlinfeng · Text-to-Video

    Camera motion synthesis system that clones diverse camera moves from multi-shot videos to animate source images, using a novel camera grid representation and a multimodal diffusion transformer. Enables aerial, dolly zoom and bullet-time movements across varied scenes without paired training data.

3D & Spatial (1)

  • Midjourney Scanner

    Midjourney · 3D Reconstruction

    Midjourney's first hardware product and new Midjourney Medical division: a full-body ultrasound device that lowers a person into a shallow pool ringed with ~500,000 ultrasonic sensors, reconstructing 3D cross-sections of muscle, fat, bone and organs ('Ultrasonic CT') in ~60s with no radiation or magnetic fields. Positioned as a faster, lower-cost MRI alternative; initially non-diagnostic body-composition maps pending FDA clearance.

Agents & Automation (5)

  • Agentic DB

    constructive-io · Agent Framework

    Postgres-based persistent memory and knowledge system for AI agents offering unified vector, BM25, full-text and spatial search across 90+ tables (conversations, memories, tasks, skills, documents, CRM). TypeScript with pgvector embeddings, PostGIS geospatial, Ollama auto-embedding — exposed via GraphQL API, generated SDK and CLI.

  • UME-Exo

    UME team · Robotics

    Universal Manipulation Exoskeleton — an upper-limb teleoperation system that delivers real-time haptic torque feedback while recording whole-arm configurations and joint-torque signals for robot policy training. Captures force-torque data alongside kinematics to teach contact-rich tasks like box flipping and GPU retrieval in constrained spaces.

  • Codex Record & Replay

    OpenAI · Coding Agents

    Codex feature that lets a user demonstrate a workflow once on their Mac, after which Codex observes the actions and window content to learn it and generates a reusable skill package replayable with different inputs via Computer Use and other tools.

  • AI Chemist (GPT-5.4)

    OpenAI · Agent Orchestration

    OpenAI paired GPT-5.4 with Molecule.one's Maria agent to improve Chan-Lam coupling of primary sulfonamides, proposing TEMPO as an additive and running 10,080 wet-lab reactions; mean yield rose 16.6%→25.2%, >30%-yield reactions rose 15.6%→37.5%, with human bench repeats confirming gains for 11 of 14 substrate pairs. Targets a bottleneck for a pharmacophore in 91+ FDA-approved drugs.

  • Moya

    DroidUp · Robotics

    Biomimetic humanoid robot (1.65 m, 32 kg) from DroidUp (Shanghai) aimed at social roles in healthcare and education, with embodied AI for real-time perception and natural movement and a claimed 92% human-like gait accuracy. Most technical specs (DOF, actuators, battery) remain unpublished.

Developer Tools (10)

  • Headroom

    chopratejas · Inference Engine

    Context compression layer for LLMs that reduces input tokens 60-95% on tool outputs, logs, RAG chunks and conversation history using algorithm-specific compressors (SmartCrusher for JSON, AST-based CodeCompressor for code, Kompress-v2 ML model for prose) with reversible compression. Ships as Python/Rust library, proxy, or MCP server; 47-92% token savings at ±0% accuracy on GSM8K/TruthfulQA/SQuAD/BFCL.

  • optimizerDuck

    itsfatduck · AI Infrastructure

    Open-source Windows optimization utility built with WPF on .NET 10, offering 30+ tweaks across performance, privacy, GPU, power and UX categories plus startup management, disk cleanup and bloatware removal with automatic backup/revert. Single portable executable, zero telemetry.

  • Gorest Spritesheet Generator

    NO6KIKO · Inference Engine

    Local browser workspace for building side-scrolling 2D game scenes — generate or import animated spritesheets, composite sprites onto backgrounds and preview animations. TypeScript, React/Vite, Express, using an AI coding assistant (Codex) for scene config and spritesheet generation.

  • AI-ML Companion

    genieincodebottle · AI Infrastructure

    AI/ML learning platform with 19 interactive tracks (foundations, linear algebra, deep learning, LLMs, MLOps, SQL) and 260+ hands-on modules featuring live visualizations, quizzes and in-browser Python execution. Bundles 10 end-to-end projects across classical ML, RAG, multi-agent orchestration and content moderation.

  • PixlStash

    Pikselkroken · AI Infrastructure

    Self-hosted image management server (Python backend, Vue frontend) that organizes large photo collections with AI tagging, auto-captioning, face detection and smart filtering via a REST API and local web UI. Supports ComfyUI workflow integration, a plugin system, read-only share tokens and GPU acceleration (CUDA/ROCm).

  • LTX-2 Trainer

    Lightricks · AI Infrastructure

    Training package for the LTX-2 audio-video model supporting LoRA training, full fine-tuning, and a flexible conditioning framework across text-to-video, image-to-video, video/audio extension, inpainting and outpainting modalities.

  • Open Knowledge Format (OKF)

    Google Cloud · AI Infrastructure

    Open, vendor-neutral specification for representing organizational knowledge (schemas, metrics, runbooks, APIs) as portable bundles of markdown files with YAML frontmatter, linked into a knowledge graph so AI agents can share context across systems without lock-in. No runtime, SDK or proprietary compression required.

  • Karpathy's LLM Wiki

    Andrej Karpathy · AI Infrastructure

    Gist describing an architecture where an LLM incrementally compiles and maintains a persistent markdown wiki from source documents — extracting insights, building cross-references and flagging contradictions once — rather than retrieving raw chunks per query like traditional RAG.

  • PixelRAG

    StarTrail-org · Data Integration

    Retrieval-augmented generation system that renders documents as screenshots instead of parsing text, enabling visual search over tables, charts and layouts via a LoRA-fine-tuned Qwen3-VL-Embedding model. Ships with a pre-built index of 8.28M Wikipedia pages.

  • Unsloth GLM-5.2 Guide

    Unsloth · Edge Deployment

    Guide for running and fine-tuning Z.ai's GLM-5.2 (744B params, 1M context) locally via Unsloth dynamic-quantized GGUFs — covers hardware requirements, KV-cache quantization and benchmarking; dynamic quantization cuts model size up to 86% while retaining reasoning, coding and agentic performance near Claude Opus / GPT-5.5 / Gemini 3.1 Pro.

Research & Safety (2)

  • iFixAi

    ifixai-ai · Alignment & Safety

    Diagnostic tool that audits AI agents for operational misalignment by running 45 automated inspections (32 core + 13 extended) across five risk categories (fabrication, manipulation, deception, unpredictability, opacity), returning a letter grade in under 5 minutes. Python 3.10+, supports OpenAI/Anthropic/Gemini/Azure/Bedrock/HuggingFace as system-under-test and cross-provider judges.

  • LOGOS

    LOGOS-Hub · Research Paper

    First multi-domain generative foundation model for the natural sciences built on a unified scientific grammar, operating as a single autoregressive transformer that tokenizes proteins, molecules, materials and reactions into a shared vocabulary without explicit 3D geometry networks.

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