Quanta AI Labs

18 / Apr 27 – May 3, 2026 · 62 RELEASES

AI News Week 18, 2026

Frontier wave: xAI Grok 4.3 (1M context, AAII 53), Mistral Medium 3.5 (128B open-weight, 77.6% SWE-Bench Verified) with Vibe Remote Agents, NVIDIA Nemotron 3 Nano Omni (30B-A3B Mamba-Transformer, 262K), SenseTime SenseNova U1 (Apache 2.0 unified multimodal), inclusionAI Ling-2.6-flash (104B MoE), and Tuna-2 (pixel embeddings beat vision encoders). Anthropic ships Claude for Creative Work — 9 connectors for Adobe CC, Blender, Resolume, Ableton. Adobe extends Firefly Video and AI color grading in Premiere. Nucleus Image (17B sparse MoE Apache 2.0), LumiVid + LumiPic HDR, Vista4D video reshooting, MoCapAnything V2, and OmniShotCut push image/video. Karpathy's autoresearch lets agents iterate nanochat training overnight; gstack, gpt-engineer revival, Conducty, jcode, Semble (98% fewer tokens than grep+read), Context Mode, AgentFM (P2P AI compute), Google's official skills repo, and a wave of self-evolving agents (Hermes, Evolver, GenericAgent) round out an enormous agent week.

Language Models (5)

  • Ling-2.6-flash

    inclusionAI · MoE Architecture

    Open-source 104B-parameter MoE instruct model with 7.4B active params, tuned for agent workloads. Hybrid linear architecture hits ~340 tokens/s and is engineered for token-efficient, production-scale automated tasks.

  • SenseNova U1

    SenseTime · Reasoning

    Native unified multimodal model series with NEO-Unify architecture eliminating visual encoder/VAE. Open-source SoTA across understanding, reasoning, and generation; 8B dense and A3B MoE variants released under Apache 2.0.

  • Nemotron 3 Nano Omni

    NVIDIA · MoE Architecture

    30B-A3B hybrid Mamba-Transformer MoE omni model unifying video, audio, image and text reasoning. 262K context, leads OCRBenchV2/WorldSense, up to 9.2x throughput vs open omni rivals. Open weights, datasets and recipes.

  • Talkie

    Talkie LM · Small Language Models

    13B 'vintage' language model trained exclusively on pre-1931 public-domain books, newspapers, journals, and case law. Contamination-free by design, enabling clean experiments on emergent capabilities and historical forecasting.

  • Grok 4.3

    xAI · Reasoning

    xAI's reasoning-always-on flagship with 1M-token context, native video input, and 207 tok/s output. ~40% input price cut to $1.25/$2.50 per Mtok; jumps to 53 on Artificial Analysis Intelligence Index.

Vision & Image (6)

  • Nucleus Image

    Nucleus AI · Text-to-Image

    Text-to-image model using sparse Mixture-of-Experts (17B total / 2B active), Apache 2.0 licensed — scales image generation efficiently by only activating a fraction of experts per token.

  • SketchVLM

    SketchVLM Team · Visual Understanding

    Lets vision-language models like GPT-4 and Gemini visually explain their answers by drawing editable, non-destructive SVG overlays — circles, arrows, labels — directly on the input image.

  • Z-Anime

    SeeSee21 · Text-to-Image

    Full fine-tune of Alibaba's Z-Image Base (S3-DiT, 6B params) for anime-style text-to-image generation. Ships multiple variants tuned for different speed/quality tradeoffs.

  • Tuna-2

    Tuna AI · Visual Understanding

    Unified multimodal model that performs visual understanding and generation directly on pixel embeddings, skipping pretrained vision encoders. Shows end-to-end pixel-space learning can match or beat encoder pipelines.

  • LumiPic

    Oumad / Lightricks · Image Editing

    Single-image SDR-to-HDR reconstruction via LogC3-encoded LoRA adapters on frozen Qwen-Image-Edit-2511 and FLUX.2-klein DiTs. Outputs scene-linear EXR with up to ~55x SDR brightness; 88-563MB LoRAs, base-model agnostic.

  • A.I.Warper 360 Equirectangular Trick

    A.I.Warper · Image Editing

    Community technique using GPT Image 2 to convert any generated scene into a 2:1 equirectangular 360 panorama by re-feeding the image with a projection prompt, producing VR-ready panoramas without dedicated panorama models.

Video & Animation (8)

  • Adobe Firefly Video + Premiere AI Color

    Adobe · Video Editing

    Adobe expands Firefly with new AI-powered video creation tools and reinvents color grading in Premiere Pro with intelligent automation, extending its lead in pro creative tooling.

  • CityRAG

    CityRAG Team · Text-to-Video

    Generates extended city street videos from a starting image, a user-defined path, and geographic data — letting you virtually navigate real urban environments with consistent weather and lighting.

  • Seedance 2.0 Skill

    dexhunter · Text-to-Video

    Agent skill for writing effective video-generation prompts for ByteDance's Jimeng Seedance 2.0. Covers input constraints, @-reference syntax, camera language, and templates for ads, dramas, MVs, and education.

  • OmniShotCut

    UVA Computer Vision Lab · Video Understanding

    Holistic shot-boundary detection model that finds cuts and classifies transition types and inter-shot relationships in one pass. Uses a shot-query Transformer trained entirely on synthetic data, no manual annotation.

  • Happy Horse

    Happy Horse · Text-to-Video

    Simple all-in-one AI video platform that consolidates the common video-creation workflow into a single tool, aimed at users who want polished output without juggling multiple specialised apps.

  • MoCapAnything V2

    AnimotionLab · Video Editing

    Second-generation motion-capture system that extracts character motion from arbitrary input sources, extending the original MoCapAnything pipeline for more robust, general-purpose mocap from in-the-wild footage.

  • Vista4D

    Eyeline Labs · Text-to-Video

    Video-reshooting framework that synthesises a dynamic scene from new camera trajectories and viewpoints using 4D point clouds, letting users refilm an existing video from angles the original camera never captured.

  • LumiVid

    Lightricks / Tel Aviv University · Video Editing

    HDR video generation method that adapts the LTX-2 video DiT with a <1% LoRA via LogC3 latent alignment. Recovers highlight/shadow detail lost in SDR and outputs scene-linear float16 EXR; VAE and base remain frozen.

Audio & Speech (3)

  • Telnyx Voice AI

    Telnyx · Voice Agents

    Voice AI agent platform that owns the full stack from carrier network to GPU inference, eliminating telco middlemen and the latency that plagues most conversational AI deployments.

  • Vobiz

    Vobiz · Voice Agents

    India-focused SIP trunk, voice API, and programmable telephony platform built specifically for embedding AI voice agents into business phone workflows.

  • MiMo V2.5 ASR

    Xiaomi · Speech-to-Text

    8B parameter end-to-end speech recognition model handling Mandarin dialects (Wu, Cantonese, Hokkien, Sichuanese), Chinese-English code-switching, song lyrics, and noisy multi-speaker audio — MIT licensed.

3D & Spatial (3)

  • Gaussian Wrapping

    Ecole Polytechnique · 3D Reconstruction

    Reconstructs watertight, textured surface meshes — including thin structures like bicycle spokes — by treating 3D Gaussians as oriented surface elements rather than blobs.

  • LaviGen

    LaviGen Team · 3D Generation

    Repurposes a 3D generative model to autoregressively place objects one at a time in a room, respecting geometric relations and physical constraints to produce realistic 3D layouts.

  • AnyRecon

    Yutian10 · 3D Reconstruction

    3D reconstruction framework that builds large scenes from sparse, unordered images using a video diffusion model to synthesize consistent novel views, supporting flexible input counts from casual captures.

Agents & Automation (18)

  • Hermes Agent

    Nous Research · Agent Framework

    Self-improving agent with a built-in learning loop — creates skills from experience, searches its own past conversations, runs across Telegram/Discord/Slack, and works with any model provider.

  • Evolver

    EvoMap · Agent Framework

    GEP-powered self-evolution engine that turns ad hoc prompt tweaks into auditable Genes and Capsules; lifts base models from 9.1% to 18.57% on CritPt.

  • GenericAgent

    lsdefine · Agent Framework

    Minimal 3K-line self-evolving agent: 9 atomic tools plus a 100-line loop give any LLM full system control over browser, terminal, filesystem, and mobile — and crystallizes each task into a reusable skill.

  • Vibe-Trading

    HKUDS · Agent Framework

    One-command personal trading agent with 74 skills, 29 swarm presets, 27 tools, and 6 data sources — turns any LLM into a market analyst, executor, and risk manager.

  • Space Agent

    Agent Zero · Agent Framework

    Agent that reshapes its own interface — ask for a page, tool, widget, or workflow and it builds it directly into the running workspace while you work.

  • gstack

    Garry Tan · Agent Framework

    Garry Tan's exact Claude Code setup — 23 opinionated specialists (CEO, designer, eng manager, QA lead, security officer) and 8 power tools as slash commands, MIT licensed.

  • gpt-engineer

    Anton Osika · Agent Framework

    OG natural-language-to-code CLI from the creator of Lovable — specify software in plain English, watch the AI write and execute the code, then iterate via follow-up requests.

  • TradingAgents

    TauricResearch · Agent Framework

    Multi-agent LLM framework for financial trading with structured-output Research Manager, Trader, and Portfolio Manager roles. Supports GPT-5, Gemini 3, Claude 4, DeepSeek, Qwen, and GLM via LangGraph.

  • AI-Trader

    HKUDS · Agent Framework

    Agent-native trading platform where any AI agent (Claude Code, Codex, Cursor, etc.) can register and trade autonomously. Includes Polymarket paper trading, market intel jobs, and a unified dashboard.

  • Google Agent Skills

    Google · Agent Framework

    Official Google repo of Agent Skills for Google Cloud and other Google products, covering Gemini API, AlloyDB, BigQuery, Cloud Run, Cloud SQL, Firebase, GKE, and Well-Architected Framework recipes.

  • autoresearch

    Karpathy · Coding Agents

    Karpathy's experiment giving an AI agent a single-GPU nanochat training setup and letting it edit train.py, run 5-min trials, and iterate autonomously overnight. The human only edits the program.md prompt.

  • Conducty

    Robert Barclay · Agent Orchestration

    Systems-level orchestration for Claude Code backed by an Obsidian vault. Replaces the plan-prompt-wait-fix loop with batch planning, parallel subagents, and a wikilinked knowledge graph that compounds across sessions.

  • RecursiveMAS

    RecursiveMAS · Agent Framework

    Recursive multi-agent framework where agents collaborate by passing latent representations through a RecursiveLink module instead of text, letting the system iteratively refine reasoning across rounds.

  • ARA: Agent-Native Research Artifacts

    Orchestra Research · Agent Framework

    Framework that restructures research papers into four machine-readable layers (cognitive, physical, exploration, evidence) so AI agents can execute, reproduce, and extend the full branching research process.

  • Moonlake 3D Agent

    Moonlake AI · Desktop Agents

    World-modeling agent with computer-use capabilities that drives Blender directly to build articulated 3D assets and large editable scenes from a single image, refining outputs via layered fidelity, consistency, and structural objectives.

  • Mistral Medium 3.5 + Vibe Remote Agents

    Mistral AI · Coding Agents

    Dense 128B model with 256K context and configurable reasoning effort scoring 77.6% on SWE-Bench Verified, released open-weight under modified MIT. Paired with Vibe Remote Agents: cloud sandboxes for parallel async coding tasks.

  • AgentFM

    AgentFM · Agent Framework

    SETI@Home for AI: peer-to-peer libp2p mesh that turns idle CPUs/GPUs into a decentralized supercomputer. OpenAI-compatible API, Podman-sandboxed agents, hardware-aware routing, end-to-end encrypted, Apache 2.0.

  • jcode

    1jehuang · Coding Agents

    Next-gen coding agent harness optimized for multi-session workflows, customizability, and resource efficiency. Boots in tens of MB RAM versus hundreds for Codex CLI, OpenCode, and GitHub Copilot CLI.

Developer Tools (18)

  • Four-Block Prompting System for Seedance 2.0

    TechHalla · Prompt Engineering

    X thread from TechHalla breaking down a four-block prompting framework that unlocks higher-quality outputs from ByteDance's Seedance 2.0 video model.

  • Metaharness

    Superagentic AI · AI Infrastructure

    Open-source Python library that optimizes the executable harness around coding agents — instructions, setup, validation, and tests — instead of just tweaking prompts.

  • NotebookLM-py

    Teng Lin · AI Infrastructure

    Unofficial Python API, CLI, and agent skill for Google NotebookLM that exposes features the web UI doesn't — bulk source import, audio/video/slides/quiz generation, and batch downloads.

  • Andrej Karpathy Skills

    Forrest Chang · Prompt Engineering

    A single CLAUDE.md file distilling Karpathy's observations on LLM coding pitfalls into four principles: think before coding, simplicity first, surgical changes, and goal-driven execution.

  • Higgsfield MCP

    Higgsfield · AI Infrastructure

    MCP server that exposes 30+ Higgsfield image and video generation models to AI agents like Claude, letting them produce cinematic media end-to-end from chat.

  • MetaTrader MCP Server

    Aria Dewangga · AI Infrastructure

    MCP bridge that lets AI assistants like Claude and ChatGPT place trades, manage positions, and stream quotes from MetaTrader 5 using natural language.

  • InsForge

    InsForge · AI Infrastructure

    Apache-licensed backend platform built for AI coding agents — exposes database, auth, storage, and functions through a semantic layer agents can reason about and operate end-to-end.

  • Understand Anything

    Lum1104 · AI Infrastructure

    Turn any codebase, knowledge base, or docs into an interactive knowledge graph you can explore, search, and query. Works with Claude Code, Codex, Cursor, Copilot, and Gemini CLI.

  • Claude Ads

    AgriciDaniel · Prompt Engineering

    Comprehensive paid-advertising audit and optimization skill for Claude Code with 250+ checks across Google, Meta, YouTube, LinkedIn, TikTok, Microsoft, and Apple Ads, plus weighted scoring and AI creative generation.

  • Toprank

    NoWork Studio · Prompt Engineering

    Open-source Claude Code skills for SEO, GEO, and Google/Meta Ads. Connects agents to Search Console and ad accounts to find wasted spend, diagnose ranking drops, and ship fixes directly to the repo.

  • Context Mode

    mksglu · AI Infrastructure

    MCP server that sandboxes tool output to keep raw data out of the context window, cutting usage by 98%. Tracks sessions in SQLite with FTS5 search so agents survive compaction across 14 platforms.

  • camofox-browser

    jo-inc · AI Infrastructure

    Anti-detection headless browser for AI agents built on Camoufox, a Firefox fork that spoofs fingerprints at the C++ level. Drop-in Puppeteer/Playwright replacement that bypasses Cloudflare and most bot detection.

  • Semble

    MinishLab · AI Infrastructure

    Fast, accurate code search library for agents that returns exact snippets using ~98% fewer tokens than grep+read. Indexes a full repo in under a second on CPU with no API keys, exposed via MCP.

  • PairDrop

    schlagmichdoch · AI Infrastructure

    Cross-platform local file sharing in the browser, inspired by Apple AirDrop and forked from Snapdrop. Sends images, documents, and text peer-to-peer between devices on the same network with no setup or signup.

  • Dictionary of AI Coding

    Matt Pocock · Prompt Engineering

    Plain-English glossary of AI coding jargon covering models, tokens, context, inference, and pricing. Aimed at demystifying the vocabulary VC marketing keeps opaque so devs can reason about cost and behavior.

  • Claude for Creative Work

    Anthropic · AI Infrastructure

    Suite of nine Claude connectors for Adobe CC, Blender, Autodesk Fusion, SketchUp, Affinity, Ableton, Splice, and Resolume. Claude acts as on-demand tutor, writes scripts, and bridges workflows; Blender connector ships via open MCP.

  • AI Engineering Newsletter

    Sumanth P · Prompt Engineering

    Weekly newsletter with hands-on tutorials for building LLM apps and AI agents. Covers OpenClaw and Claude Code, Gemma 4 fine-tuning, RAG architectures, voice cloning, agent memory, and code-review automation.

  • Modal + OpenAI Agents SDK Webinar

    Modal Labs · AI Infrastructure

    Webinar with Modal's Erik Dunteman on building with OpenAI's new open-source Agents SDK and Modal Sandboxes. Modal is an official sandbox provider, giving agents GPU-capable computers for code, research, and tool-use workflows.

Research & Safety (1)

  • Heretic

    p-e-w · Alignment & Safety

    Fully automatic censorship removal for transformer LLMs using directional ablation plus an Optuna optimizer — produces decensored models that match human-expert abliterations with much lower KL divergence.

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