Quanta AI Labs

20 / May 11 – May 17, 2026 · 54 RELEASES

AI News Week 20, 2026

Neural Computers week — AI agents moving into ambient personal compute. Claude Code ships goal-oriented loops + Karpathy's CLAUDE.md discipline. OpenAI launches Codex from anywhere and ChatGPT personal finance. Google DeepMind reimagines the mouse pointer for AI interaction. Krea 2 ships first foundation model with stylistic control. Qwen 3.6 27B MTP, MiniCPM-V 4.6, and Cactus Needle (26M function-call model for tiny devices) advance frontier and edge inference. dflash speculative decoding. Agent ecosystem explosion: agency-agents (full AI agency), hyperresearch (agent-driven research wiki), dograh open-source voice platform, alumnium AI E2E testing, claude-context MCP, MinerU document processing, Google's official Skills repo. Immersive wave: AIDC Pixelle-Video automated short-video engine, Supertonic on-device multilingual TTS, ResembleAI Dramabox, Scenema voice cloning, SeeDance2Stitcher (30s seamless videos), Articraft 3D assets, TrackCraft3R dense 3D tracking, Pixal3D, NVIDIA SANA-WM minute-scale world model, PhyMotion physics-grounded human video, Khala music AI. Research: SIGMAP explainable receptor affinity, MAMMAL biomedical foundation model (npj Drug Discovery), Thinking Machines on Interaction Models, Warp-as-History. Tooling: Hermes desktop companion, openhuman personal AI, streambert media app, logto SaaS auth.

Language Models (2)

  • Qwen3.6-27B-MTP-Q4_K_M-GGUF

    RDson · Text Generation

    Quantized GGUF format version of Qwen 3.6 27B model with Q4_K_M quantization, enabling efficient local deployment of the language model with reduced memory footprint.

  • Codex API Access

    OpenAI · Code Generation

    API endpoint enabling programmatic access to OpenAI's Codex code generation model from any application or environment. Supports code completion, synthesis, and transformation tasks via REST interface.

Vision & Image (3)

  • Krea 2 Launch Announcement

    Krea AI · Text-to-Image

    Krea ships its first foundation model, built from scratch for aesthetic diversity and stylistic control. Positions against generic text-to-image with editorial-grade style fidelity.

  • AsymFlow

    Stanford University (Hansheng Chen et al.) · Text-to-Image

    Rank-asymmetric flow model for pixel-space image generation achieving 1.57 FID on ImageNet 256×256. Enables finetuning latent models (FLUX.2 klein) into pixel-space models while beating latent baselines on HPSv3, DPG-Bench, and GenEval benchmarks.

  • MiniCPM-V-4.6

    OpenBMB · Visual Understanding

    A multimodal vision-language model combining compact architecture with visual understanding capabilities. Designed for efficient inference while maintaining performance on vision-language tasks.

Video & Animation (11)

  • Gemini Omni (Leaked)

    Google (via TestingCatalog) · Text-to-Video

    Evidence of an upcoming Gemini Omni video model spotted in the Gemini mobile app: "Meet our new video model. Remix your videos, edit directly in chat, try a template, and more." Expected at Google I/O 2026.

  • Gemini Omni vs Seedance 2.0 Comparison

    Chetaslua · Text-to-Video

    Side-by-side video output comparison between the leaked Gemini Omni model and ByteDance Seedance 2.0, evaluating motion coherence, prompt adherence, and visual quality.

  • Dreamina Octo

    ByteDance Dreamina · Text-to-Video

    New creative system from Dreamina built around Seedance 2.0, positioned as a creative partner that understands intent (not just a prompt-responder). Coming soon.

  • MoCam

    Research · Text-to-Video

    Camera motion-controllable video generation system for cinematic shot synthesis with explicit camera path control.

  • Relit-LiVE

    Research · Video Editing

    Relightable live video editing system for production-grade lighting transfer and post-capture relighting of dynamic footage.

  • CausalCine

    Research · Text-to-Video

    Real-time autoregressive generation for multi-shot video narratives — generates coherent shot sequences with causal scene transitions for film-like outputs.

  • Seed Dance 2.0

    ByteDance · Text-to-Video

    Unable to extract - page content could not be fetched. Based on URL pattern, likely a ByteDance AI/ML project or framework update.

  • Pixelle-Video

    AIDC-AI · Text-to-Video

    Video generation or processing framework leveraging the Pixelle architecture. Likely focused on efficient video synthesis or manipulation with neural rendering capabilities.

  • SeeDance2Stitcher

    TomClive · Video Editing

    Unable to determine - page content could not be fetched from GitHub repository

  • Sana-WM

    NVIDIA Labs · Text-to-Video

    Efficient world model for minute-scale video generation and prediction, enabling real-time synthesis of long-horizon visual sequences with reduced computational requirements.

  • PhysMotion

    UCLA, University of Utah · Text-to-Video

    Generates physics-grounded video dynamics from a single image by combining 3D Gaussian splatting with differentiable Material Point Method simulation and diffusion-based refinement. Produces physically plausible motion constrained by continuum mechanics priors rather than pure data-driven generation.

Audio & Speech (4)

  • Supertonic

    Supertone Inc. · Text-to-Speech

    Unable to determine - GitHub page content could not be fetched. Based on company background, likely a voice synthesis or audio AI tool.

  • Dramabox

    ResembleAI · Text-to-Speech

    Voice generation and synthesis model for dramatic character audio. Designed for creating expressive speech with theatrical inflection and emotional variation.

  • Scenema Audio

    Scenema · Audio Processing

    Audio generation or processing tool from Scenema, likely for creating or manipulating audio content, though specific technical details are unavailable due to page fetch limitations.

  • Khala

    Khala-Music-AI · Music Generation

    Music generation and processing AI system designed for audio synthesis and music creation tasks. Limited details available from repository metadata.

3D & Spatial (5)

  • Viggle PINOC

    Viggle AI · 3D Generation

    Free mocap tool: upload a motion video, get skeleton animation as .fbx/.glb ready for Blender or Maya. Bonus: upload a character image, get 3D Gaussian splatting model with motion preview. Replaces thousand-dollar mocap pipelines.

  • Pixa3D

    ldyang694 · 3D Generation

    3D object generation and reconstruction system from images using neural rendering. Generates high-quality 3D models with texture and geometry from 2D inputs.

  • Viggle AI

    Viggle Inc. · 3D Generation

    Motion capture and 3D generation platform that extracts skeletal animation from video without mocap suits, generates 3D characters and scenes, and enables real-time character swapping for meme creation and game content. Includes PINOC for motion extraction and Fight Anyone 3D, a PvP fighting game with AI-generated 3D motions.

  • TrackCraft3r

    CVLAB KAIST · 3D Reconstruction

    3D object tracking and reconstruction system that combines tracking with 3D shape estimation from video sequences. Enables real-time multi-object tracking with 3D geometry recovery.

  • Articraft

    Articraft3D · 3D Generation

    An agentic system for automated generation of articulated 3D assets at scale, combining AI agents with 3D modeling to produce rigged and animated character models.

Agents & Automation (3)

  • DreamX-World

    Research · Agent Framework

    Generative world model for interactive scene exploration and long-horizon environment simulation.

  • OpenHuman

    tinyhumansai · Agent Framework

    Unable to determine - page content could not be fetched from GitHub repository.

  • Personal Finance ChatGPT

    OpenAI · Web Agents

    ChatGPT application focused on personal finance tasks, enabling users to get financial guidance and analysis through conversational AI.

Developer Tools (17)

  • Claude Code Stocks + Crypto MCP

    TwelveData (via gus on X) · Data Integration

    Single-command MCP install gives Claude Code live access to 17,000+ stocks, crypto prices, and financial statements in seconds. Demonstrated by gus showing one MCP add command turning Claude Code into a real-time financial analyst.

  • Anime Fight Storyboard Prompt

    Dheepan Ratnam · Prompt Engineering

    Reference prompt structure for generating a 15-second 16:9 anime fight previsualization with shot-by-shot timing (0-3s, 3-6s, etc.), camera angles, and action beats — useful template for video model prompting.

  • Seedance 2.0 Valence-Arousal Prompting

    Kōda · Prompt Engineering

    Experiments using emotional valence-arousal coordinates as prompting axes for Seedance 2.0 (via Mitte AI) — likely transferable to other video/image models for affective scene control.

  • Seedance 2.0 30s Seamless Stitcher

    TomLikesRobots · AI Infrastructure

    Tool for chaining Seedance 2.0 video extensions without the overlapping-frame and brightness-shift artifacts at join points. Detects duplicate frames, trims overlap, matches tone for clean 30s+ outputs.

  • Karpathy's 4 CLAUDE.md Rules

    Andrej Karpathy · Prompt Engineering

    Karpathy shares 4 CLAUDE.md rules that reduced Claude Code mistakes from 41% to 11% across 30 codebases — focused on planning discipline, constraint stating, and explicit no-go zones.

  • MCP

    TwelveData · Data Integration

    Model Context Protocol implementation for financial data integration, enabling AI agents to access real-time market data and trading information through standardized protocol interfaces.

  • dflash

    z-lab · Inference Engine

    Unable to determine - repository page could not be fetched. Based on naming convention, likely a flash attention or efficient transformer inference optimization library.

  • SIGMAP

    alberdom88 · AI Infrastructure

    Unable to determine - page content could not be fetched from GitHub repository.

  • Hermes Desktop

    fathah · Inference Engine

    Desktop application for running Hermes LLM locally with a graphical interface. Enables local inference of the Hermes language model without cloud dependencies.

  • Needle

    Cactus Compute · Evaluation & Benchmarks

    Unable to determine - page content could not be fetched from GitHub repository

  • Alumnium

    alumnium-hq · AI Infrastructure

    Unable to determine - page content could not be fetched from GitHub repository

  • MinerU

    OpenDataLab · Data Integration

    Document understanding and conversion tool that extracts structured data from PDFs and other document formats using machine learning. Designed for high-fidelity conversion of complex documents including layouts, tables, and embedded content.

  • Logto

    logto-io · AI Infrastructure

    Open-source identity and access management (IAM) platform providing authentication, authorization, and user management infrastructure with support for multiple protocols and integrations.

  • StreamBERT

    truelockmc · Inference Engine

    Unable to determine - GitHub page content could not be fetched. Likely a BERT-based streaming or real-time processing framework based on naming convention.

  • Claude Code Common Workflows

    Anthropic · Prompt Engineering

    Documentation for Claude Code integration patterns and common use cases. Provides guidance on implementing code generation and execution workflows with Claude API.

  • Filmustage

    Filmustage · AI Infrastructure

    AI-powered screenwriting and film production tool that automates script breakdown, scheduling, and budgeting for filmmakers and production teams.

  • KREA 2

    KREA · AI Infrastructure

    Unable to determine - page content could not be fetched

Research & Safety (9)

  • Gemini 3.2 Release Prediction

    Polymarket · Survey & Overview

    Prediction market projects 88% chance Gemini 3.2 ships within one week — a signal of strong insider belief in an imminent Google frontier model release.

  • The Bystander Effect in Multi-Agent Reasoning

    arXiv · Research Paper

    Paper quantifying "cognitive loafing" in collaborative multi-agent LLM interactions — agents perform worse on tasks when other agents are present, mirroring the human bystander effect. Implications for agent team design.

  • Alex Cooper AI Industry Post

    Alex Cooper (LinkedIn) · Survey & Overview

    AI industry commentary post from Alex Cooper on LinkedIn covering recent model releases and platform shifts.

  • AI Practitioner LinkedIn Post

    LinkedIn · Survey & Overview

    Notable AI practitioner post shared on LinkedIn covering frontier model developments and practical agent deployment lessons.

  • HyperResearch

    jordan-gibbs · Research Paper

    Unable to determine - GitHub page content could not be fetched. Appears to be a research-related project based on repository name.

  • CS224N: Natural Language Processing with Deep Learning

    Stanford University · Survey & Overview

    Graduate-level course covering neural network fundamentals for NLP and large language models, with lectures, PyTorch assignments, and final projects. Covers deep learning basics through cutting-edge LLM research with publicly available videos and materials.

  • Mouse Pointer Interaction Redesign

    Google DeepMind · Research Paper

    Research on reimagining human-AI interaction paradigms through pointer-based interfaces, exploring new modalities for AI system control and navigation.

  • MAMMAL

    IBM Research · Research Paper

    Molecular Aligned Multi-Modal Architecture and Language — a foundation model for biomedical discovery published in npj Drug Discovery, unifying small molecules, proteins, and natural language for drug-target interaction prediction.

  • Warp-as-History

    yyfz · Research Paper

    Unable to determine - page content could not be fetched from repository

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