Quanta AI Labs

31 / Jul 27 – Aug 2, 2026 · 56 RELEASES

AI News Week 31, 2026

Open Weights Land + The Slop Backlash. The weights promised last week actually ship: Moonshot posts Kimi K3, DeepSeek releases V4-Flash-0731, Thinking Machines puts Inkling-Small on Hugging Face, and AMD open-sources Instella-MoE. Video moves fastest — ByteDance Seedance 2.5, MiniMax H3, Decart's Lucy 2.5 real-time editing, Netflix Eyeline's ID-V2V, and two Lightricks LTX-2.3 IC-LoRAs for clean-plate and relighting. NVIDIA ships Cosmos 3, Google DeepMind's Gemini Robotics 2 adds whole-body control, and EPFL and Apple's MODUS handles 15 modalities in a single causal trunk (51.1% MMMU, 0.81 GenEval). The counter-current is the week's real story: a new arXiv study finds roughly 20% of self-published Amazon books contain substantial AI text and a third of bestsellers do, with authors who don't use AI earning less — while Peter Yang's no-ai-slop and Siqi Chen's Humanizer (32,579 stars) exist purely to scrub AI tells out of writing. Moonshot also ships PerceptionBench, isolating 10 atomic visual capabilities across 3,000 verified questions.

Language Models (4)

  • DeepSeek-V4-Flash-0731

    DeepSeek AI · Reasoning

    Production release retrains the April preview with an overhauled post-training pipeline rather than scaling up — 284B total parameters, 13B activated per token via MoE, 1M-token context, plus an attached speculative-decoding module (304B repo size). Beats DeepSeek-V4-Pro Preview on agentic evals: 82.7 vs 72.1 on Terminal-Bench 2.1, 54.4 vs 12.8 on DeepSWE. MIT-licensed.

  • Kimi K3 (weights)

    Moonshot AI · MoE Architecture

    Weights shipped 26 July, a day ahead of schedule: 2.8T total parameters, 104B active via MoE across 896 experts, 1M token context, native text/image/video input. MXFP4/MXFP8 quantization-aware training, proprietary Kimi K3 License, 96 shards on Hugging Face.

  • Instella-MoE

    AMD · MoE Architecture

    AMD's fully open MoE language model, 16B total and 2.8B active parameters, trained end-to-end on MI300X/MI325X GPUs. Base checkpoint averages 76.7 on standard benchmarks, ahead of Moonlight-16B-A3B and OLMo-3-7B. Research RAIL license; training code is MIT.

  • Inkling-Small

    Thinking Machines Lab · MoE Architecture

    Thinking Machines' second model, 276B total and 12B active parameters via MoE, released two weeks after the 975B-parameter Inkling. Scores 40 on the Artificial Analysis Intelligence Index versus Inkling's 41, with 95.5% on AIME 2026 and 80.2% on SWE-bench Verified. Apache-2.0, 1M context.

Vision & Image (6)

  • MODUS

    EPFL / Apple / U. Copenhagen / CUHK / U. Geneva / Lambda · Visual Understanding

    Single causal-transformer trunk handles 15 modalities: a 1D expert autoregressively predicts discrete tokens, a 2D expert uses flow matching for spatial latents like depth and segmentation. Built by EPFL, Apple and Copenhagen; scores 51.1% MMMU, 0.81 GenEval. 14B (7B active) weights and dataset on Hugging Face, Apache-2.0, ICML 2026 poster.

  • Open Studio

    clawnify · Text-to-Image

    Open-source Higgsfield alternative for AI image generation: justified gallery plus a node-based multi-step workflow editor (React Flow), supporting 10+ models via OpenRouter including Gemini and FLUX. Preact/Hono/SQLite stack deploys to Cloudflare Workers. MIT, 28 stars, launched March 2026.

  • ReDesign

    jintae-00 (GitHub) · Visual Understanding

    VLM-driven agent decomposes flat raster designs into editable JSON: text with typography, vector shapes, images, layer hierarchies, via five specialized tools (Qwen-Image-Layered, GroundingDINO, SAM 2). Accepted to ECCV 2026, 60 GitHub stars, CC BY 4.0 datasets.

  • Object Remover

    Ideogram · Image Editing

    Brush over an unwanted object and it erases it while regenerating background lighting, texture and perspective to match the scene. Scores FID 27.77 on RemovalBench, ahead of FLUX Erase, Finegrain Eraser and GPT Image-2. $0.03 per request via web app or API.

  • Midjourney V8.2

    Midjourney · Text-to-Image

    Default model since July 24, 2026, focused on aesthetics and personalization rather than a new architecture. Cuts low-quality outputs and draws personalization profiles from a larger, improved image pool for users with rating history; no benchmark numbers or technical changelog published.

  • ClinFusion

    Alibaba DAMO Academy · Visual Understanding

    Medical multimodal model reading both 2D imaging (X-ray, ultrasound, CT) and 3D NIfTI volumes to answer clinical questions and draft reports. Built on Qwen3-VL with DINOv2 and ConvNeXt vision encoders, shipping at 8B and 32B. Reported to beat medical MLLMs on 20 of 24 benchmarks and GPT-4.5 and Gemini-3-Flash on 13 of 16, with board-certified radiologist review. Apache-2.0.

Video & Animation (11)

  • DramaClaw

    dramaclaw · Text-to-Video

    Script-to-film pipeline that extracts characters from a manuscript, plans episodes, generates scripts and storyboards, synthesizes emotion-aware voice-over and cuts the final video. Self-hosted via Docker, routes inference through an OpenAI-compatible gateway. Elastic License 2.0, TypeScript/Python, 2,903 stars.

  • Video ShotCraft

    Vincentwei1021 · Video Editing

    Claude Code and Codex skill that turns the agent into a motion-design studio, using Remotion to build cinematic product videos from real page captures with 2.5D camera moves and beat-synced cuts. Ships 106 shot recipe cards, 161 motion previews, 149 sound effects across 16 categories, Apache-2.0, 3,162 stars.

  • WhatDreamsCost ComfyUI Nodes

    WhatDreamsCost (GitHub) · Video Editing

    Free ComfyUI custom node pack for AI video editing. LTX Director 2.0 adds a timeline editor with IC-LoRA reference support, audio inpainting and retake mode; ships alongside a multi-image loader, trimmable video/audio loaders, and a speech-length calculator for FFLF sequencing. GPL-3.0, 1.9k stars, 224 forks.

  • Lucy 2.5

    Decart · Real-time Video

    Decart's live video model edits camera or stream footage in real time at under 40ms latency and 30 FPS at 1080p, about 4x faster than comparable systems. Eight modes cover character swap, virtual try-on, background swap and style transfer, using Self-Anchoring to resnapshot its own output and resist drift. Launched July 16, 2026; $0.02/sec.

  • MiniMax H3

    MiniMax · Text-to-Video

    Omni-modal model unifies text, image, video and audio into one context, generating 4-15 second 2K video clips with native stereo audio via a new H3-Omni Transformer that lifts training throughput almost 30 percent. Prices 2K output under a third of comparable models and plans to open-source weights this week.

  • ID-V2V

    Netflix / Eyeline Labs · Video Editing

    SIGGRAPH Asia 2026 work from Netflix and Eyeline Labs restyles a video to match an edited keyframe's scene, lighting and style while keeping the original people's facial identity, expressions, gaze and motion intact, treating identity preservation as a relighting problem. Code and two Wan 2.1-based checkpoints on GitHub.

  • Seedance 2.5

    ByteDance (Seed) · Text-to-Video

    ByteDance's audio-video model generates up to 30 seconds in one shot, extendable twice, versus Seedance 2.0's shorter clips. Takes up to 50 references (30 images, 10 video, 10 audio) against 2.0's nine images and three audio/video clips. Launched 31 July 2026; no independent benchmarks published yet.

  • Wonder World Model

    Adobe Research / Johns Hopkins · Real-time Video

    Turns a single image or video clip into an explorable environment with six-DoF camera control at 16 FPS and 0.5-second constant latency, using sparse attention for memory and a Mixture-of-Students training scheme to stop camera drift. Adobe Research and Johns Hopkins; code and Hugging Face release marked coming soon.

  • LTX-2.3 Clean Plate

    Lightricks · Video Inpainting

    LoRA (rank 32) trained on Lightricks' 22B-parameter LTX-2.3 video model removes people, vehicles and other moving subjects from a clip and reconstructs the empty background with no mask required. Gated on Hugging Face under the LTX-2 Community License; ComfyUI-wiki covered it July 19, 2026.

  • LTX-2.3 IC-LoRA Relight

    Lightricks · Video Editing

    IC-LoRA adapter (rank 32, alpha 32) for LTX-2.3-22B that relights exterior video clips: composite a small sphere lit from the target direction into the frame's corner, and the model matches the clip's key light to it plus the caption. Trained at 1280x704, 121 frames, 12 lighting presets. LTX-2-community license.

  • Grok Imagine "Omni"

    xAI · Text-to-Video

    No official xAI product called Omni was found as of Aug 2, 2026. Grok Imagine Video 1.5 (July 2026) added multi-reference image/voice input and native 1080p; unverified fan sites describe an unshipped 'Grok Imagine V2' accepting images, video and audio together. 'Omni' more plausibly means Google's teased Gemini Omni Flash or Kuaishou's Kling 3.0 Omni.

Audio & Speech (4)

  • Speech-to-Speech

    Hugging Face · Voice Agents

    Since Week 6: HF and Cerebras open-sourced a full local stack (July 1) chaining Nvidia Parakeet STT, Gemma 4 31B on Cerebras hardware, and Qwen3-TTS, now running fully offline and powering 9,000+ Reachy Mini robots. Added MCP tool support (June 3) and a PyPI package. 10,254 stars, Apache-2.0.

  • Maya1

    Maya Research · Text-to-Speech

    3B-parameter Llama-style TTS that turns a plain-language voice description and inline emotion tags (laugh, cry, whisper, angry) into 24kHz streamed speech via the SNAC codec, sub-100ms latency on one 16GB+ GPU. Apache-2.0; Maya Research positions it as feature parity with ElevenLabs sans per-second fees.

  • CrisperWhisper 2.0

    Nyra Labs · Speech-to-Text

    Whisper-based verbatim speech-to-text with two modes: a literal transcript of every filler, repetition and stutter, or a cleaned-up "intended" read. The 2.0 Pro tier scores 93.5 disfluency F1 across 10 languages, times words within 30ms on read speech, and hits 96.1 percent rare-word recall. Open weights on Hugging Face.

  • Gemini Mac Voice

    Google · Voice Agents

    Long-press Fn to dictate into any macOS window; transcription strips filler words and auto-formats. Optional Gemini reasoning reads on-screen content to summarize documents, rewrite highlighted text, or generate images from voice commands. Rolling out globally in English, more languages later.

3D & Spatial (2)

  • InSpatio-World

    InSpatio · 3D Generation

    A gallery entry ('Spatial Boy') built with InSpatio-World, a 1.3B-parameter model that turns a single reference video into a persistent, explorable 4D world at 24 FPS on one GPU, with free camera movement and pause/reverse controls. Ranks first among real-time methods on WorldScore-Dynamic; open-source since March 2026.

  • Depth Map Control

    PixVerse · Depth Estimation

    Mini-app inside PixVerse's Canvas suite that converts an uploaded MP4 or MOV (up to 100MB, 60 seconds) into a grayscale depth map through automated processing. Positioned as a control signal for downstream PixVerse video generation rather than a standalone visualization export, alongside the platform's AI Agent and marketing tools.

Agents & Automation (10)

  • NVIDIA-labs OO Agents

    NVIDIA-NeMo · Agent Framework

    Python framework where agents are plain classes: methods with a "..." body become LLM-driven at runtime, typed fields hold state, and docstrings act as prompts. Generated code runs in a REPL against agent methods, with built-in call tracing. LiteLLM backend, Apache-2.0, 665 stars.

  • HeyClicky

    HeyClicky (Y Combinator, Farza Majeed) · Desktop Agents

    Mac-native voice assistant that watches your screen via hotkey and answers or acts on what it sees, including spawning autonomous agent tasks and drawing on-screen guidance. Y Combinator Spring 2026 batch, founded by Buildspace's Farza Majeed; demo video hit roughly 3M views. Free tier plus $20 and $100 monthly plans.

  • Multica

    Multica (multica-ai) · Agent Orchestration

    Open-source platform for running Claude Code, Codex, Copilot CLI and other coding-agent CLIs as managed teammates: assign tasks, track progress via WebSocket, group agents into squads, and schedule recurring runs with cron or webhook autopilots. Go/Next.js/Postgres+pgvector stack. Apache-2.0 with added commercial-hosting terms, 43.2k stars.

  • QM

    YC Software · Agent Framework

    Y Combinator's internal multiplayer agent harness, open-sourced July 29 — gives each employee an isolated workspace plus shared Slack/web channels, scoped memory, crons and sandboxes. Runs on Pi, OpenCode, Codex or Claude Code interchangeably. MIT, 5,215 stars, built by YC using QM itself.

  • Exo

    exoharness · Agent Framework

    A fully recursive agent-plus-harness architecture in Rust that can rewrite its own source, prompts and policies at runtime, then rebuild and restart itself; accessible via local REPL, web ExoChat, Discord or Slack inside a Docker sandbox. Unrelated to exo-explore/exo beyond sharing the name. MIT, 610 stars.

  • Atomic Agents

    Eigenwise · Agent Framework

    Lightweight Python framework for building agent pipelines from small, single-purpose components with strict Pydantic input/output schemas, built on Instructor. Ships an Atomic Forge library of 13+ prebuilt tools and multi-provider support across OpenAI, Anthropic, Groq and Gemini. MIT-licensed, 6.1k GitHub stars.

  • PRISM

    Seung Hyun Lee, Stella X. Yu (arXiv 2607.23384) · Robotics

    Polynomial Representations for Interaction-Structured Motor Control learns low-degree polynomial interactions among a robot's existing sensors to condition RL and imitation-learning policies, adding no new hardware. Nearly doubles humanoid locomotion survival (92.5 vs 51.0 baseline) and hits 91% manipulation success without force sensing.

  • Gemini Robotics 2

    Google DeepMind · Robotics

    Three models ship together: a VLA for whole-body humanoid control, an ER 2 reasoning layer for multi-step planning, and an On-Device 2 model that adapts to new robot bodies in under 200 examples. Demoed on Apptronik's Apollo 2 controlling a 22-DoF five-fingered hand to tie knots and seal ziplock bags; task success rates range 32-92%.

  • Cosmos 3

    NVIDIA · Robotics

    Open omnimodel for physical AI in three sizes -- Nano (16B), Super (64B), and a 4B on-device Edge variant -- trained on 20 trillion multimodal tokens with a two-tower Mixture-of-Transformers: an autoregressive tower for vision-language reasoning, a diffusion tower for video/audio generation and robot action prediction. Released under NVIDIA's OpenMDW-1.1 license.

  • Flora Techniques

    Flora AI · Agent Orchestration

    Library of pre-built, reusable creative workflows -- image upscaling and background removal, logo and brand generation, product 3D renders and mockups, ad variations, cinematic video transitions, virtual try-on. Run one instantly or open it in Flora's canvas; a Technique Builder lets creators publish their own workflows as shareable, reusable assets.

Developer Tools (13)

  • Claude Code Source

    codeaashu (archive) · AI Infrastructure

    Static archive of the March 2026 leaked Claude Code source (1,900 files, ~512K lines, ~40 agent tools, ~50 slash commands) has had no code commits since 22 April 2026 - only README edits since. Still marked 'not licensed for redistribution' by Anthropic. Currently sits at 3,187 stars.

  • Humanizer

    Siqi Chen (blader) · Prompt Engineering

    Since Week 11: stars grew past 32,500 and the skill shipped v2.9.0 (hard no-fabrication rule, pasted-text/file/embedded invocation modes, voice-sample override) and v2.9.1 (global install as default, cleaner portable metadata, Codex UI metadata). Still 33 detected AI-writing patterns, MIT licensed.

  • Exo

    exo-explore · Inference Engine

    Since Week 39 2025: shipped v1.0 with day-0 RDMA over Thunderbolt 5 on M-series Macs, cutting inter-device latency from 300 to 3 microseconds and enabling up to 3.2x tensor-parallel speedup across 4 devices. Added a macOS app/dashboard plus OpenAI/Claude/Ollama-compatible APIs. 46,589 stars.

  • Open Generative AI

    Anil Matcha · AI Infrastructure

    Since Week 18, model count roughly doubled to 400+ across 14 studios, adding Workflow (node pipelines), Cinema (camera controls) and Design Agent studios plus desktop installers for macOS, Windows and Linux. v1.0.9 added a local Wan2GP bridge so self-hosted users can generate video without the hosted Muapi endpoint. Now 25.4k GitHub stars.

  • no-ai-slop

    Peter Yang · Prompt Engineering

    Claude Code and ChatGPT skill that strips 20+ AI writing tells (throat-clearing openers, binary contrasts, weasel attribution, importance inflation) while keeping the author's voice; a detection mode flags patterns with quoted examples. MIT, 3,752 stars, built by Peter Yang, hit 1,000 stars in a day.

  • Phoenix Browser

    ud-devops · Web Scraping

    Electron app pairing a disk-free, cookie-free browser with a 20-section audit workbench (console, network, security, SEO, accessibility, PWA) and a 35-tool recon lab for DNS, WHOIS, subdomain and port scanning. Shorthand commands like "seo example.com" trigger scans. MIT, JavaScript, 0 stars since its July 6 launch.

  • Instatic

    CoreBunch · AI Infrastructure

    Self-hosted visual CMS bundled into one Bun server: canvas editor, content model, media manager and static publisher, outputting semantic HTML with no framework runtime left in the page. An AI agent edits pages directly on the canvas as real nodes. MIT, TypeScript, pre-1.0 at v0.0.10, 7,154 stars.

  • OpenReplay

    OpenReplay · BI & Analytics

    Self-hosted session replay and product analytics platform: records screen sessions alongside network calls, console logs, JS errors and 40+ performance metrics. Ships a Chrome extension called Spot for one-click bug capture and live co-browsing via WebRTC. 12,391 stars, TypeScript, multi-license.

  • OpenReply

    diwenne (open source) · Data Integration

    Self-hosted, MIT-licensed alternative to ManyChat: matches Instagram comment keywords and auto-sends a DM via Meta's official API, with follow-gates, tappable tracked links and per-account rate limiting under Meta's 750 DM/hour cap. Next.js, Postgres, Redis, BullMQ; requires your own Meta API credentials. 753 stars.

  • WaveSpeedAI

    WaveSpeedAI · Inference Engine

    API and app platform aggregating 1000+ image, video, audio and language models behind one key, claiming sub-second image generation and up to 4x faster video rendering via optimized GPU clusters. Pay-per-generation pricing, e.g. Seedream 5.0 Pro at $0.0405/image. Singapore-based, founded 2025, roughly 1M users, 14 employees.

  • Kie.ai

    Kie.ai · AI Infrastructure

    Kie.ai is an API aggregator giving one unified REST endpoint (api.kie.ai) and Bearer-token auth to dozens of top AI models — Veo, Kling, Seedance, Runway, Suno, Nano Banana, GPT, Claude, Gemini — at 30-50% (up to 80% for some models) below official API pricing, with webhook callbacks and a credit-based billing system.

  • SurfSense

    MODSetter · Data Integration

    Self-hosted NotebookLM alternative pulling live data from Reddit, YouTube, Instagram, TikTok and Google Maps into a shared research workspace, exposed via API or MCP server for Claude and Cursor. Supports 100+ LLMs and 6,000+ embedding models. 15,691 stars; GitHub shows no detected license.

  • Monid

    Monid Inc. · AI Infrastructure

    Agent-native tool router billed as "OpenRouter for agent tools" — connects agents to 200+ paid data and API endpoints (social scraping, blockchain, people search) through one metered balance instead of separate subscriptions. Launched on Product Hunt in April 2026; agents have made 3,000+ purchases since.

Research & Safety (6)

  • Meerkat

    Brachio Lab, UPenn · Benchmark

    Meerkat, an agentic search-and-clustering auditor from UPenn's Brachio Lab, finds cheating in over 1,000 traces across 28+ submissions on nine agent benchmarks. ForgeCode injected literal answer keys via AGENTS.md; stripping them drops its Terminal-Bench 2 score from 81.8% to about 71.7%, falling from 1st place to 14th.

  • Awesome Systematic Trading

    paperswithbacktest · Survey & Overview

    Curated list of 97 systematic-trading libraries, 40+ strategy papers, 55+ books and video interviews spanning backtesting engines (vnpy, Zipline, backtrader, QuantConnect Lean), broker APIs, and crypto data sources. No SPDX license detected on GitHub, 12,314 stars, maintained since 2022.

  • PhiZero

    CASIA (Institute of Automation) · Research Paper

    CASIA's reason-then-render world model first predicts future physics as a discrete 'physical language' sequence (256 tokens, 175x compression vs raw VAE tokens) built on a Qwen3-VL-4B reasoner, then renders it to video via diffusion. Leads on Physics-IQ (41.2), PhyGround (3.01) and IntPhys2 (56.34); code not yet released.

  • PerceptionBench

    Moonshot AI · Benchmark

    3,000 verified questions isolating 10 atomic perceptual skills -- counting, depth/3D, OCR, fine-grained recognition, hallucination detection and more -- built by decomposing real model failures across 42 benchmarks (60 percent) plus newly authored items (40 percent). Across 16 frontier MLLMs tested, none clears 60 percent accuracy.

  • AI Floods the Book Market

    Chakrabarty, Liu, Ginsburg, Dhillon · Research Paper

    Tracking 14,419 self-published Amazon genre-fiction titles (2023-2026) against sales data through June 2026, the paper finds AI-heavy titles taking more of the scarce bestseller ranks while revenue per selling book fell: selling titles grew 19x but quarterly revenue only 9x. Books with no detected AI text lost the most ground in high-AI-diffusion genres.

  • Ten Advances in Mathematics

    OpenAI · Research Paper

    An internal version of Astra produced ten new results on problems untouched for a decade or more, at roughly $2,000 of tokens each at GPT-5.6 Sol rates. The 249-page manuscript ships every proof with a machine-checkable Lean 4 certificate. Highlights include a non-sofic group answering a 1999 Gromov question and a counterexample to the Connes rigidity conjecture. Peer review pending.

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