Quanta AI Labs

16 / Apr 13 – Apr 19, 2026 · 33 RELEASES

AI News Week 16, 2026

NVIDIA Lyra 2.0 and World Labs Spark 2.0 push 3D world generation toward production with single-image scene reconstruction and browser-native Gaussian splatting. Baidu open-sources ERNIE-Image (8B DiT) for dense-text image generation. HeyGen launches HyperFrames (HTML→MP4) and Zed ships Zeta 2 code edit prediction. Claude Code ecosystem explodes with awesome-lists, MCP servers, and SEO/research/DevOps agents. ACE-Step v1.5 Turbo and MOSS-TTS-Nano advance commercial-grade audio. reverse-SynthID raises watermarking robustness concerns.

Language Models (3)

  • Gemma-4-31B-it int4 AutoRound

    Intel · Small Language Models

    Int4-quantized variant of Google's Gemma 4 31B instruction-tuned model produced via Intel's AutoRound (signed gradient descent weight rounding), shrinking the model to ~6B effective size while preserving quality for efficient inference.

  • Ternary Bonsai

    Prism ML · Small Language Models

    Collection of 1.58-bit ternary-quantized Bonsai LLMs (1.7B, 4B, 8B) released in MLX 2-bit and GGUF formats, dramatically reducing memory footprint while preserving text-generation quality for on-device inference.

  • Zeta 2

    Zed Industries · Code Generation

    8B-parameter code edit prediction model fine-tuned from ByteDance Seed-Coder-8B-Base that predicts next-edit suggestions for cursor regions in the Zed editor, using SPM-style prompts with git-merge markers to define editable code.

Vision & Image (1)

  • ERNIE-Image

    Baidu · Text-to-Image

    8B-parameter open-source single-stream Diffusion Transformer text-to-image model with exceptional dense-text rendering, structured generation (posters, comics), and complex instruction following. Includes a Turbo variant (8 steps via DMD+RL) that runs on 24GB consumer GPUs.

Video & Animation (1)

  • HappyHorse 1.0

    HappyHorse AI · Text-to-Video

    No-signup browser-based AI video generator producing 1080p HD multi-shot videos from text or image inputs in ~10 seconds with a 99.5% success rate. Supports photorealism, anime, cyberpunk, watercolor styles and aspect ratios from 16:9 to 21:9.

Audio & Speech (3)

  • MOSS-TTS-Nano

    OpenMOSS · Text-to-Speech

    0.1B-parameter open-source TTS model that generates 48kHz stereo audio in 20 languages with voice cloning, designed for CPU-only inference without GPUs. Pure autoregressive architecture with streaming generation and ~2x efficiency via ONNX CPU backend.

  • ACE-Step v1.5 XL-Turbo

    ACE Studio & StepFun · Music Generation

    4B-parameter text-to-music DiT model that generates high-quality audio in just 8 distilled inference steps (no CFG). Trained on legally-compliant licensed/royalty-free/synthetic data, making it commercial-ready and runnable on 12GB+ VRAM.

  • VoiceMode

    mbailey · Voice Agents

    MCP server enabling natural two-way voice conversations with Claude Code and other agents. Supports local Whisper.cpp + Kokoro TTS for offline use or OpenAI Whisper API, with intelligent silence detection for hands-free interaction.

3D & Spatial (3)

  • Lyra 2.0

    NVIDIA Spatial Intelligence Lab · 3D Generation

    Framework for generating persistent, explorable 3D worlds from a single image. Combines long-horizon video generation with feed-forward 3D reconstruction, addressing spatial forgetting and temporal drifting via per-frame geometry routing. 14B-parameter model outputs Gaussian Splats and meshes for Isaac Sim.

  • Spark 2.0

    World Labs · NeRF & Gaussian

    Web-based 3D Gaussian Splatting renderer integrating with THREE.js and WebGL2. Adds a level-of-detail system, .RAD progressive streaming format, GPU virtual memory (16M splats with 64K chunk paging), and multi-object compositing for browser-based 3DGS on phones and VR headsets.

  • WonderZoom

    Stanford University · 3D Generation

    Multi-scale 3D world generation system that lets users zoom into regions and generate detailed content from text prompts while maintaining cross-scale geometric consistency via depth registration and scale-adaptive Gaussian surfels.

Agents & Automation (3)

  • Verticals v3 (YouTube Shorts Pipeline)

    rushindrasinha · Agent Orchestration

    Open-source AI content engine that turns a topic into a published YouTube Short via a research → script → visuals → voiceover → captions → assembly → upload pipeline. Supports Claude/Gemini/GPT/Ollama, Edge TTS/ElevenLabs/Kokoro, and 15 niche profiles.

  • Feynman

    Companion AI · Agent Framework

    Open-source AI research agent (TypeScript CLI) that automates literature review, paper analysis, and experiment replication via specialized multi-agent workflows. Slash commands like /deepresearch, /lit, /replicate; integrates AlphaXiv, Modal/RunPod GPUs, and local models.

  • Siclaw

    Scitix · Coding Agents

    Open-source DevOps/SRE AI agent for read-only infrastructure diagnostics. Performs a 4-phase evidence-gathering and root-cause workflow with investigation memory, reusable runbooks, and MCP tool extensions, accessible via terminal, web UI, or chat.

Developer Tools (13)

  • HyperFrames

    HeyGen · AI Infrastructure

    Open-source (Apache 2.0) video rendering framework that turns HTML compositions into deterministic MP4 videos using headless Chrome and FFmpeg. Designed agent-first with a CLI, 50+ ready-made blocks, and Frame Adapters for GSAP, Lottie, Three.js, and CSS animations.

  • Atropos

    Nous Research · Evaluation & Benchmarks

    Async-first LLM RL environments framework for collecting and evaluating model trajectories across diverse environments (GSM8K, MMLU, multi-turn games, code execution, multimodal). Demonstrates 4.6x improvement on parallel function calling and integrates with Axolotl, Tinker, vLLM, and SGLang.

  • Meta-Harness

    Stanford IRIS Lab · AI Infrastructure

    Framework for automated search and optimization of task-specific model harnesses — the supporting code around a fixed base model that decides what to store, retrieve, and show. Reference implementations cover memory-system optimization and Terminal-Bench 2.0 scaffold evolution.

  • Mesh LLM

    Mesh-LLM · Inference Engine

    Distributed inference system that pools GPU capacity across networked machines behind a single OpenAI-compatible API. Auto-selects pipeline parallelism for dense models or expert sharding (zero cross-node traffic) for MoE, with inter-model collaboration during inference.

  • Claude SEO

    Daniel Agrici · Prompt Engineering

    Comprehensive Claude Code skill bundle providing 20 SEO sub-skills including technical audits, schema markup, AI search (GEO) optimization, E-E-A-T quality assessment, and integrations with Google Search Console, PageSpeed, and GA4 APIs.

  • Kaggle MCP for Claude Code

    Composio · AI Infrastructure

    MCP integration that connects Claude Code to Kaggle, enabling natural-language commands for downloading competition data, managing dataset versions, and submitting entries via Composio's tool router (20,000+ tools across 1,000+ apps).

  • Gymnasium v1.3

    Farama Foundation · AI Infrastructure

    Maintained successor to OpenAI Gym for reinforcement learning environments released v1.3.0, providing the standard RL API across Classic Control, Box2D, MuJoCo, Atari, and third-party environments.

  • agtop

    ldegio · AI Infrastructure

    Top-style terminal dashboard for monitoring AI coding agents. Auto-discovers Claude Code and Codex sessions, tracks spend with plan-aware billing, monitors context-window pressure (CTX%), and displays live CPU/memory and tool-activity history.

  • wacli

    Peter Steinberger · AI Infrastructure

    Go-based WhatsApp CLI built on whatsmeow that syncs message history to local SQLite with full-text search, sends messages and media, manages contacts/groups, and exposes presence and diagnostics — useful as an MCP-style tool surface for agents.

  • n8n-MCP

    Romuald Czlonkowski · AI Infrastructure

    MCP server that exposes n8n's automation platform to AI assistants, covering 1,505 nodes (812 core + 693 community) with 99% property schema coverage, 2,646 example configurations, and 2,709 workflow templates for natural-language workflow building.

  • Antigravity Awesome Skills

    sickn33 · AI Infrastructure

    Installable library of 1,435+ reusable SKILL.md playbooks for AI coding agents covering development, testing, security, infrastructure, product, and marketing. Works across Antigravity IDE, Claude Code, Cursor, Codex CLI, and Gemini CLI.

  • Skylos

    duriantaco · AI Infrastructure

    Open-source static analyzer for Python, TypeScript, and Go that detects unused code, hardcoded secrets, SAST vulnerabilities (SQLi, XSS, SSRF), and AI-generated code flaws like hallucinated function calls. Framework-aware (Django, Flask, FastAPI, Next.js) with VS Code, GitHub Actions, and MCP integrations.

  • CodeBurn

    AgentSeal · AI Infrastructure

    Terminal observability tool that tracks AI coding-assistant spend across Claude Code, Codex, Cursor, OpenCode, Pi, OMP, and Copilot by reading session data from disk (no API keys). Includes one-shot success tracking, an Optimize command for waste patterns, model comparison, and a macOS menubar app.

Research & Safety (6)

  • Gemma-4 E4B OBLITERATED

    OBLITERATUS · Red Teaming

    Abliterated/uncensored version of Google's Gemma 4 E4B model with guardrails surgically removed via the OBLITERATUS aggressive method (whitened SVD + attention head surgery), achieving 0% hard refusal rate. Released for research, red-teaming, and creative use.

  • reverse-SynthID

    aloshdenny · Red Teaming

    Research project that reverse-engineers Google's SynthID watermarking system using spectral analysis, achieving 90% detection accuracy and a multi-stage attack (V4) that defeats SynthID watermarks on Gemini 3.1 Flash Image and Nano Banana Pro outputs.

  • California Creative Job Losses Aren't AI Casualties

    The Hollywood Reporter · Survey & Overview

    Otis College report finds California's creative economy lost 114,000 jobs (14%) from 2022-2025, but AI-exposed roles (writers, developers, artists) actually grew. Losses stem from post-Peak-TV cost displacement, not AI; AI is replacing tasks, not staffers.

  • Awesome Claude Plugins

    Composio · Survey & Overview

    Curated awesome-list of production-ready Claude Code plugins organized into 9 categories: integrations, frontend & design, git, code quality, backend, DevOps, docs & security, productivity, and personality.

  • Claude Code Ultimate Guide

    Florian Bruniaux · Survey & Overview

    24,000+ line educational resource for Claude Code with 247 production-ready templates (agents, commands, hooks, skills), 41 Mermaid diagrams, a 7-module learning path, 271-question quiz, and a security threat database tracking 24 CVEs and 655 malicious skill patterns.

  • Claude Code Projects Index

    Daniel Rosehill · Survey & Overview

    Curated index of 75+ Claude Code agent workspaces and 28 cluster plugins spanning sysadmin, legal, health, finance, content, and research — demonstrating Claude Code as a general-purpose agent workspace beyond just coding.

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