Quanta AI Labs

29 / Jul 13 – Jul 19, 2026 · 35 RELEASES

AI News Week 29, 2026

Open-Weights Frontier + Motion/Video Surge. Frontier drops cluster: Moonshot Kimi K3 (2.8T-param open MoE, 1M context), Thinking Machines Lab ships Inkling — its first open-weights frontier model (975B/41B MoE, 45T multimodal tokens) — xAI open-sources Grok Build (Rust coding-agent TUI), Prism ML Bonsai 27B (ultra-low-bit ternary/1-bit quantization), and NVIDIA Nemotron-3-Embed tops the RTEB embedding benchmark. Image: NVIDIA/Comfy-Org PixelDiT (VAE-free pixel-space diffusion, CVPR Best Paper Finalist), ByteDance Seedream 5.0 Pro now API-available on BytePlus, and Lucida matting. Video + human animation: Qualcomm MobileWan (on-device 5B video gen on phones), Alibaba Wan-Dancer (minute-long music-to-dance) and Wan-Streamer v0.3 (duplex audio-video), MeiGen-AI talking-head hub, and the Cardboard AI editor. Motion + world/3D: RobbyAnt LingBot World 2 (now on Reactor), Motion4Motion cross-morphology transfer, Snap SnapMoGen, NVIDIA ARDY (humanoid motion to Unitree G1), Google DeepMind GenCeption (video models as vision learners), and Google GNM 3D head model. Big agent-tooling + Agent-Skills wave: Colibri (run 744B GLM-5.2 on 25GB RAM), Google Always-On Memory Agent, Graphify code knowledge graphs, official GSAP Skills, taste-skill anti-slop UI, Super Video Maker, scroll-world, nano-banana image skills, Digital Marketing Pro, and a Telnyx/Kimi voice agent. Safety: CopySight IP-compliance. Industry: George Lucas embraces AI (Variety).

Language Models (3)

  • Kimi K3

    Moonshot AI · MoE Architecture

    2.8-trillion-parameter open MoE model (16 of 896 experts active via Stable LatentMoE, Kimi Delta Attention) with native multimodal vision and a 1M-token context, scoring 67.3 on DeepSWE; priced at $0.30/MTok cached input and $15/MTok output.

  • Bonsai 27B

    Prism ML · Small Language Models

    Ultra-low-bit quantized multimodal LLM based on Qwen3.6 27B, with a ternary {-1,0,+1} variant (1.71 bits/weight, 5.9 GB, ~95% of full-precision performance) and a 1-bit binary variant (1.125 bits/weight, 3.9 GB, ~90% retention). Quantizes end-to-end with no higher-precision escape hatches, keeping 262K context, tool calling, vision, and speculative decoding.

  • Inkling

    Thinking Machines Lab · MoE Architecture

    First open-weights frontier model from Mira Murati's Thinking Machines Lab: a multimodal Mixture-of-Experts LLM (975B total / 41B active, up to 1M-token context) trained on 45T multimodal tokens across text, image, and audio, with controllable thinking effort, uncertainty calibration, and agentic coding/tool use. Weights on Hugging Face, fine-tuning via Tinker.

Vision & Image (4)

  • Seedream 5.0 Pro

    ByteDance · Text-to-Image

    ByteDance multimodal image model with deep-thinking prompt reasoning, real-time web search, precision layer editing to alpha PNGs, native 2K output, and on-image text in 10+ languages; outputs feed the Seedance 2.x image-to-video pipeline. Now available via API on BytePlus.

  • nano-banana-skills

    ShinChven · Image Editing

    MIT-licensed collection of AI agent skills for Nano Banana Pro (Google Gemini) covering image transformations like anime-to-photorealistic, character reference sheets, cosplay conversion, IMAX 70mm portrait outpainting, and 8K photo restoration; used by adding the repo as Knowledge to a Gemini Gem.

  • Lucida

    Ege Orcun · Image Editing

    Image matting / background-removal model fine-tuned from BiRefNet (MIT) for hard cases like semi-transparent objects, camouflage, typography, glows, and illustrations, outputting a 1024x1024 alpha channel. Trained on 52,882 image-alpha pairs; reports 0.0257 MAE on a 203-image benchmark, distributed via Hugging Face transformers.

  • PixelDiT

    NVIDIA (NVlabs) / Comfy-Org · Text-to-Image

    Single-stage, end-to-end pixel-space diffusion transformer that generates high-res images directly with no VAE, using a dual-level design (patch-level DiT for global semantics + pixel-level DiT for texture). Achieves 1.61 FID on ImageNet 256 and 0.74 GenEval; the Comfy-Org repo ships ~61GB of ComfyUI-ready safetensors incl. 512→2048 and 1024→4096 upscalers. (CVPR 2026 Best Paper Finalist.)

Video & Animation (8)

  • Super Video Maker

    Distribb (Bomx) · Video Editing

    Open-source end-to-end AI video-production Agent Skill that lets agents produce polished videos: HeyGen avatar explainers, FFmpeg screen-recording/demo composition, B-roll via Seedance 2.0 and OpenAI images, Whisper captioning, and Remotion motion graphics, plus layout/QC checks.

  • scroll-world

    oso95 · Video Editing

    Agent skill for Claude Code/Codex that generates scroll-scrubbed fly-through landing pages where a camera continuously flies through interconnected 3D diorama scenes, using a vanilla-JS scrub engine plus ffmpeg and orchestrating Higgsfield (GPT Image 2) for stills and Seedance/Kling for frame-locked transitions.

  • Cardboard

    Cardboard · Video Editing

    Browser-based AI video editor that auto-finds best moments, builds edits from conversational prompts, and reframes aspect ratios, generating AI captions/voiceovers and rendering ~1 hour of footage in ~5 minutes; from $32/month.

  • MobileWan

    Qualcomm AI Research · Text-to-Video

    On-device video diffusion transformer adapting Wan2.2-5B to run a 5B-parameter model on commercial mobile hardware, generating 5-second 480x832 videos at 16 FPS in ~20 seconds (VBench 83.79) via recurrent distillation into a chunk-wise autoregressive process, causal linear attention, and learnable head pruning (23% fewer params).

  • Wan-Dancer

    Tongyi Lab (Alibaba) · Text-to-Video

    14B diffusion-based music-to-dance video system that generates stable 720p/30fps human dance clips exceeding one minute across five genres, using a two-stage hierarchical framework (global keyframe planning + local refinement), time-mapped RoPE for audio-visual alignment, optical-flow continuity loss, and LoRA choreography customization.

  • Motion4Motion

    Ling-Hao Chen et al. (Tsinghua/StepFun/HKUST) · Video Inpainting

    SIGGRAPH 2026 training-free method for cross-morphology motion transfer (human-to-animal, animal-to-object) that avoids skeletons by operating on dense pixel-level motion flow, rewiring cached self-attention K/V features in a frozen WAN text-to-video diffusion transformer via retargeted positional encodings (TransPE).

  • Wan-Streamer v0.3

    Alibaba (Tongyi/Wan) · Real-time Video

    Native-streaming, full-duplex audio-video world model generating synchronized speech and video in real time at 640x368/25 FPS with ~200ms model-side (~550ms total) latency. v0.3 keeps v0.2's serving topology but adds free-form behavior to the agent's event stream, interleaving natural-language actions with speech.

  • MeiGen-AI

    MeiGen-AI · Text-to-Video

    Open-source research hub for audio-driven human-animation models, notably MultiTalk (multi-person conversational video from multi-stream audio + reference images) and InfiniteTalk (unlimited-length talking/singing video from a portrait or silent video, built on Wan 2.1, up to 10 min at 480p/720p) with accurate lip-sync, head pose, and body motion.

Audio & Speech (2)

  • Customer Support Voice Agent

    Sumanth077 · Voice Agents

    Sample project building an inbound-call customer-support voice agent on the Telnyx AI Assistant Builder, using Kimi-K2.5 for inference plus Telnyx native STT/TTS. A FastAPI webhook injects live context (queue wait times, system status) at call start, with a WebRTC widget for browser testing.

  • Mirelo Audio-to-MIDI

    Mirelo AI · Audio Processing

    Decoder-only transformer that transcribes a full musical mix into per-instrument MIDI tracks by treating transcription as language modeling over 5-second mel-spectrogram chunks. Trained via synthetic pre-training (~1.45M MIDI files), fine-tuning on 11,000+ hours of real audio, then RL, with automatic instrument/chord/key detection and DAW export.

3D & Spatial (2)

  • SnapMoGen

    Snap Research · 3D Generation

    NeurIPS 2025 text-to-motion project releasing a dataset and pretrained checkpoints for generating 3D human motion from expressive text prompts, pairing a multi-scale residual motion VQ-VAE with a masked Transformer, plus HumanML3D evaluation, retargeting, and Blender visualization tooling.

  • GNM (Generative Anthropometric Model)

    Google · 3D Generation

    Open ecosystem of parametric 3D human models (Apache-2.0), shipping GNM Head — a statistical 3D head/face geometry model with disentangled control over identity, expression, pose, and internal anatomy (eyes, teeth, tongue), with NumPy/JAX/PyTorch/TensorFlow backends.

Agents & Automation (6)

  • taste-skill

    Leonxlnx · Agent Framework

    MIT-licensed "anti-slop" collection of portable Agent Skills that push AI-generated UIs toward better design, with aesthetic presets (soft, minimalist, brutalist) and dials for layout variance, motion, and density. Framework-agnostic (React/Vue/Svelte), installable via npx skills for Claude Code, Codex, and Cursor.

  • GSAP Skills

    GreenSock · Coding Agents

    GreenSock's official Agent Skills repo teaching AI coding agents to use GSAP correctly, with eight skills covering the core API, timelines, ScrollTrigger, plugins, utilities, React, performance, and Vue/Svelte. Installable via npx skills or the Claude Code plugin marketplace; works with Cursor, Copilot, and 40+ agents.

  • Grok Build

    xAI · Coding Agents

    xAI's open-sourced terminal-based coding agent, a full-screen Rust TUI (Apache 2.0) that understands codebases, edits files, runs shell commands, searches the web, and manages long-running tasks. Runs interactively, headless for CI, or embedded in editors via the Agent Client Protocol.

  • Always-On Memory Agent

    Google Cloud Platform · Agent Framework

    Gemini/ADK reference agent that runs continuous memory consolidation via three sub-agents (Ingest, Consolidate on a 30-min timer, Query), ingesting 27 file types into SQLite with entity/importance extraction and no vector DB or embeddings; built on Gemini 3.1 Flash-Lite.

  • Digital Marketing Pro

    indranilbanerjee · Agent Framework

    Open-source MIT-licensed AI marketing plugin with 158 skills and 25 specialist agents that runs a 12-part strategy flow producing ~50–60 strategy docs in about an hour; works across Claude Code, Codex, Cursor, and 40+ platforms via the Agent Skills standard, with 8 real API connectors (Slack, HubSpot, Klaviyo, etc.).

  • ARDY

    NVIDIA & ETH Zürich · Robotics

    Autoregressive diffusion model for real-time interactive 3D human motion generation from online text prompts and kinematic constraints (waypoints, keyframes, sparse joints), integrated with the SONIC tracking policy to drive a Unitree G1 humanoid robot; SIGGRAPH 2026 / ACM TOG.

Developer Tools (7)

  • Scrapling

    D4Vinci · Web Scraping

    Python 3.10+ web-scraping framework with adaptive selectors that relocate when page layouts change, plus HTTP/stealth/Playwright fetchers, TLS-fingerprint spoofing, and Cloudflare/Turnstile anti-bot bypass. Adds a Scrapy-like spider layer with concurrent crawling and checkpoint pause/resume; BSD-3-Clause.

  • Graphify

    Graphify Labs · AI Infrastructure

    MIT-licensed coding-assistant skill that turns codebases, docs, and media into queryable knowledge graphs using local tree-sitter AST parsing (no data leaves the machine) across 36+ languages plus PDFs, images, and SQL schemas. Outputs interactive graph.html/JSON with Leiden community detection and query/path/explain commands.

  • Claude Code UI Toolkit (curated list)

    Alphenex.ai · Prompt Engineering

    Publicly shared Google Doc curating a ranked list of ~10 frontend-design skills and MCP servers for building UI with Claude Code in 2026, including frontend-design, ui-ux-pro-max, shadcn-ui, GSAP, Framer Motion, and Convex, each with links.

  • Colibrì

    JustVugg · Inference Engine

    Apache-2.0 zero-dependency C runtime that runs the 744B-parameter MoE model GLM-5.2 on ~25GB-RAM consumer machines by streaming expert weights from disk across a VRAM/RAM/NVMe tier hierarchy, with MLA attention, speculative decoding, and an OpenAI-compatible API. ~0.05–0.1 tok/s on a 25GB device, up to ~6.8 tok/s with six GPUs.

  • nano-banana-pro-prompts

    YouMind-OpenLab · Prompt Engineering

    Skill (Claude/Cursor/Codex/Gemini-CLI) that recommends AI image prompts from a library of 10,000+ community-curated templates synced daily, returning up to 3 ranked prompts each with a mandatory sample image. Optimized for Nano Banana Pro but compatible with DALL-E, Midjourney, Stable Diffusion, and Flux.

  • Stock-Prediction-Models

    huseinzol05 · Inference Engine

    Archived ML/DL collection for stock forecasting bundling ~18 deep-learning price models (LSTM, GRU, Seq2seq, VAE, attention) and ~23 reinforcement-learning trading agents (Q-learning, policy gradient, actor-critic, evolution strategies), plus Monte Carlo portfolio simulations and a TensorFlow.js browser demo.

  • Nemotron-3-Embed

    NVIDIA · AI Infrastructure

    Open embedding-model family (8B and 1.14B, 4096/2048 dims, 32K context) for RAG, agentic retrieval, code search, and agent memory. The 8B variant ranks #1 on RTEB (78.5%) and is SOTA on the LMEB long-context memory benchmark (64.4); the NVFP4 1B variant delivers ~2x throughput on Blackwell.

Research & Safety (3)

  • CopySight

    CopySight AI · Alignment & Safety

    IP-compliance service for generative AI that scores AI outputs for visual similarity to protected characters, brands, and artwork, checks dataset lineage and human authorship, and logs the prompt-to-output pipeline to help register content with the U.S. Copyright Office.

  • GenCeption

    Google DeepMind · Research Paper

    ECCV 2026 paper ("Video Generation Models are General-Purpose Vision Learners") that repurposes a pre-trained video diffusion model (WAN 2.1, 14B) into a single feed-forward, text-steered vision model for depth, normals, segmentation, keypoints, camera pose, and 4D reconstruction, matching or beating task-specific SOTA with 7–500x fewer training frames.

  • George Lucas Embraces AI

    Variety · Survey & Overview

    Variety reports George Lucas voicing support for AI in filmmaking, framed around his long history with test screenings and effects innovation — industry commentary rather than a model or product release.

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