Quanta AI Labs

14 / Mar 30 – Apr 5, 2026 · 58 RELEASES

AI News Week 14, 2026

Gemma 4 and Qwen 3.6 lead frontier models alongside Qwen 3.5 Omni and GLM-5V-Turbo. Agent ecosystem explodes with Claude Code Agent Teams, gstack, Superpowers, LangGraph, and 15+ agent frameworks. Research highlights include TRELLIS 2 for 3D generation, VOID video inpainting, DreamLite on-device diffusion, and OmniVoice 600-language TTS. Context-1 and Bonsai-8B push efficient inference boundaries.

Language Models (9)

  • Context-1

    Chroma · MoE Architecture

    20B-parameter MoE model for agentic retrieval that decomposes complex queries into subqueries, iteratively searches corpora, and self-prunes irrelevant documents at 10x lower cost.

  • Kronos

    shiyu-coder · Text Generation

    Open-source foundation model for financial market prediction that tokenizes K-line candlestick data and uses autoregressive Transformers for forecasting, trained on data from 45+ global exchanges.

  • Bonsai-8B

    Prism ML · Small Language Models

    End-to-end 1-bit quantized 8B-parameter language model achieving 14x size reduction (1.15 GB) and 6x faster inference while maintaining competitive benchmark scores with full-precision models.

  • TimesFM 2.5

    Google Research · Reasoning

    200M-parameter pretrained foundation model for time-series forecasting supporting 16K context length and continuous quantile forecasting up to 1K horizons.

  • Qwen 3.6

    Alibaba / Qwen Team · Text Generation

    Latest release in Alibaba's Qwen open-weight model family with improvements to reasoning and agentic capabilities, continuing the series after Qwen 3.5.

  • Gemma 4

    Google · Text Generation

    Google's most advanced open-source language models purpose-built for advanced reasoning and agentic workflows, delivering exceptional performance-per-byte efficiency.

  • Gemma 4 E2B

    Google DeepMind · Text Generation

    2.3B effective parameter multimodal model processing text, images, and audio with 128K context window, reasoning modes, and native function-calling optimized for on-device deployment.

  • Qwen 3.5 Omni

    Alibaba / Qwen Team · Text Generation

    End-to-end omni-modal model processing and generating text, images, audio, and video with real-time streaming voice and video interaction and state-of-the-art cross-modal performance.

  • GLM-5V-Turbo

    Zhipu AI · Code Generation

    Multimodal vision-language model for vision-based coding tasks, excelling at frontend code generation from design mockups, GUI automation, and agentic workflows requiring visual understanding.

Vision & Image (5)

  • Granite 4.0 3B Vision

    IBM Research · OCR & Document

    Enterprise-grade 4B-parameter vision-language model for document data extraction including charts, tables, and key-value pairs from images into structured formats like CSV, JSON, and HTML.

  • Translumo

    ramjke · OCR & Document

    Real-time screen translator for PC games capturing on-screen text via multiple OCR engines with ML-based scoring. Supports translation to 30+ languages through DeepL, Google Translate, and others.

  • Gen-Searcher

    MMLab / UCLA / UC Berkeley · Text-to-Image

    Search-augmented image generation system performing multi-hop reasoning to collect textual knowledge and reference images for grounded generation of knowledge-intensive images.

  • See-Through

    Shitagaki Lab / UPenn / Spellbrush · Image Editing

    Automatically decomposes anime character illustrations into up to 23 semantically distinct fully inpainted layers with inferred drawing order for manipulatable 2.5D models.

  • DreamLite

    ByteDance · Text-to-Image

    Lightweight on-device unified diffusion model (0.39B parameters) performing both text-to-image generation and text-guided editing on mobile devices, generating 1024x1024 images in ~3 seconds.

Video & Animation (5)

  • OpenScreen

    siddharthvaddem · Video Editing

    Free open-source desktop app for creating product demos and screen recordings with automatic zoom effects, dual audio capture, annotations, motion blur, and variable speed control.

  • VOID

    Netflix / INSAIT · Video Inpainting

    Video object removal framework performing physically-plausible inpainting using a vision-language model and video diffusion model for consistent results in complex interaction scenarios.

  • TokenDial

    Adobe Research / CMU · Video Editing

    Converts pretrained text-to-video models into continuous video editors by learning lightweight slider controls that adjust semantic attributes like appearance and motion through additive token offsets.

  • HyDRA

    Huazhong University / Kuaishou (Kling Team) · Video Understanding

    Hybrid memory architecture for video world models maintaining visual coherence for hidden subjects during out-of-view intervals with spatiotemporal relevance-driven retrieval.

  • VGGRPO

    Google / University of Copenhagen / Oxford · Text-to-Video

    Latent geometry-guided reinforcement learning framework improving video diffusion models' geometric consistency and world coherence using 4D scene understanding and latent-space RL rewards.

Audio & Speech (2)

  • LongCat-AudioDiT

    Meituan · Text-to-Speech

    Diffusion-based TTS model operating directly in waveform latent space, achieving superior voice cloning through a simplified architecture combining waveform VAE with diffusion backbone.

  • OmniVoice

    k2-fsa · Text-to-Speech

    Massive multilingual zero-shot TTS model scaling to over 600 languages using a non-autoregressive discrete architecture that directly maps text to multi-codebook acoustic tokens.

3D & Spatial (6)

  • GaussianGPT

    Technical University of Munich · 3D Generation

    Transformer-based model that autoregressively generates 3D Gaussian scenes via next-token prediction, supporting unconditional generation, scene completion, outpainting, and controllable sampling.

  • TRELLIS 2

    Microsoft · 3D Generation

    Large-scale 3D asset generation model creating high-quality 3D assets from text or image prompts in multiple formats including Radiance Fields, 3D Gaussians, and meshes with models up to 2B parameters.

  • Generative World Renderer

    Alaya Studio · 3D Generation

    Toolkit extracting G-buffers from commercial games for generative editing, environment stylization, material prediction, and video relighting across 4M+ frames from AAA titles.

  • MC-PBR Master

    MicroCreativity · PBR Materials

    Professional ComfyUI custom node pack for end-to-end PBR texture generation, enabling generation, refinement, upscaling, and visualization of physically-based rendering maps entirely within ComfyUI.

  • LGTM

    HKU / Apple · 3D Reconstruction

    First native 4K feed-forward method predicting compact textured Gaussians for high-resolution 3D novel view synthesis, combining geometric primitives with per-primitive textures.

  • HandX

    UIUC / Snap Inc. · 3D Generation

    Large-scale dataset and generation framework for realistic bimanual hand motion with fine-grained finger dynamics via diffusion and autoregressive models.

Agents & Automation (19)

  • LangGraph

    LangChain · Agent Framework

    Low-level orchestration framework for building, managing, and deploying long-running stateful AI agents with durable execution, human oversight, and persistent memory.

  • CatchMe

    HKUDS · Coding Agents

    Lightweight vectorless personal memory system that captures your entire digital footprint locally and makes it queryable by CLI agents like Claude Code, Cursor, and others.

  • SkillRL

    Aiming Lab · Agent Framework

    Reinforcement learning framework enabling LLM agents to learn reusable behavioral skills from past experiences by abstracting successful trajectories into a hierarchical skill library.

  • oh-my-codex

    Yeachan Heo · Coding Agents

    Workflow layer for OpenAI Codex CLI that adds canonical skills like deep-interview, planning, and team coordination with persistent project state management.

  • Meta-Harness TBench2

    Stanford IRIS Lab · Coding Agents

    Terminal-Bench 2.0 competition artifact extending Terminus-KIRA with environment snapshot injection, achieving 76.4% accuracy by reducing early exploration steps in agentic coding tasks.

  • Claw Code

    Ultraworkers · Coding Agents

    Open-source autonomous coding harness built and maintained by AI agents, orchestrating parallel coding sessions, testing, documentation, and workflow automation through modular tools.

  • AI Marketing Skills

    Single Brain · Agent Framework

    Production-ready automation workflows for marketing and sales including experiment engines, lead qualification, content scoring, and SEO intelligence for Claude Code and other AI agents.

  • Paperclip

    PaperclipAI · Agent Orchestration

    Node.js server and React UI for coordinating multiple AI agents to run autonomous businesses with org charts, task management, budget enforcement, and multi-agent orchestration.

  • GoClaw

    nextlevelbuilder · Agent Framework

    Multi-tenant AI agent platform in Go supporting 20+ LLM providers, 7 messaging channels, multi-agent orchestration with task delegation, and built-in prompt injection detection.

  • GeeLark

    GeeLark · Web Agents

    AI-powered social media management platform providing cloud-based virtual phones for managing multiple accounts across mobile platforms without physical devices.

  • lil agents

    ryanstephen · Coding Agents

    Animated AI companions living on macOS dock as transparent video characters. Click to open an AI terminal supporting Claude Code, Codex, Copilot, and Gemini CLIs with thinking bubbles.

  • Claude Code Agent Teams

    Anthropic · Agent Orchestration

    Experimental feature for orchestrating multiple Claude Code instances as a coordinated team with shared task list, inter-agent messaging, and lead/teammate architecture via tmux.

  • Superpowers

    Jesse Vincent / Prime Radiant · Agent Framework

    Complete software development workflow for coding agents built on composable skills including brainstorming, TDD, planning, subagent-driven development, and git worktrees.

  • OpenSpace

    HKUDS · Agent Framework

    Self-evolving skill platform enabling AI agents to automatically create, refine, and share reusable skills across a community, reducing token consumption by 46% while improving task performance.

  • Visual Explainer

    nicobailon · Coding Agents

    Agent skill converting complex terminal output into styled HTML pages and slide decks with interactive Mermaid diagrams, diff reviews, and responsive layouts for AI coding agents.

  • CCG Workflow

    fengshao1227 · Agent Orchestration

    Multi-model orchestration system where Claude Code coordinates task routing between Gemini and Codex with 29+ slash commands, patch-based security, and parallel agent execution.

  • MiroFish Offline

    nikmcfly · Agent Framework

    Fully local multi-agent simulation engine generating hundreds of AI agents with unique personalities to simulate social media reactions, sentiment evolution, and opinion dynamics without cloud APIs.

  • AutoAgent

    kevinrgu · Agent Framework

    Autonomous agent engineering framework allowing a meta-agent to iteratively improve an agent harness overnight by modifying prompts, tools, and configuration while hill-climbing on benchmark scores.

  • PSDesigner

    Fudan University / NTU · Web Agents

    Automated graphic design system mimicking human designer workflows by autonomously collecting theme-related assets and executing design tool operations to produce professional PSD files.

Developer Tools (9)

  • Pretext

    Cheng Lou · AI Infrastructure

    Pure JavaScript/TypeScript library for measuring and laying out multiline text without DOM interactions, avoiding layout reflow by implementing its own text measurement using the browser's font engine.

  • Airbyte

    Airbyte · Data Integration

    Leading open-source data integration platform for ETL/ELT pipelines with 600+ connectors to move data from APIs, databases, and files to warehouses and lakehouses.

  • Metabase

    Metabase · BI & Analytics

    Open-source business intelligence platform that lets anyone query, visualize, and analyze data without SQL expertise, now featuring AI-powered analytics via Metabot.

  • OLMo-core

    Allen Institute for AI · Inference Engine

    PyTorch building blocks for the OLMo ecosystem providing foundational training scripts, inference utilities, and model implementations for the OLMo family of open language models.

  • Meta-Harness

    Stanford IRIS Lab · Evaluation & Benchmarks

    Optimization framework that automatically improves AI model evaluation harnesses by analyzing filesystem logs of source code, execution traces, and performance scores.

  • Awesome Design MD

    VoltAgent · Prompt Engineering

    Curated collection of 55+ DESIGN.md files from popular websites, enabling AI coding agents to generate consistent pixel-perfect UI with design tokens, color palettes, and component styles.

  • G0DM0D3

    Elder Plinius · AI Infrastructure

    Privacy-focused open-source chat interface supporting 50+ AI models via OpenRouter with parallel multi-model evaluation, input perturbation for red-teaming, and semantic output transformation.

  • Awesome Agent Skills

    VoltAgent · Prompt Engineering

    Curated collection of 1,060+ real-world agent skills from engineering teams at Anthropic, Google, Vercel, Stripe, and others for extending Claude Code, Codex, and Gemini CLI.

  • Greptile

    Tabnam Inc. · AI Infrastructure

    AI code review tool that automatically analyzes pull requests with full codebase understanding, identifies bugs and security issues, and provides inline comments with diagrams.

Research & Safety (3)

  • L1B3RT4S

    Pliny the Prompter · Red Teaming

    Collection of adversarial prompt injection and jailbreak techniques for red-teaming flagship LLMs and AI platforms, organized by target system.

  • LLM Visualization

    Brendan Bycroft · Interpretability

    Interactive 3D animated visualization that walks through the internal mechanisms of large language models, making transformer architecture visually explorable.

  • Attention Residuals

    Kimi Team · Research Paper

    Research paper proposing AttnRes, a technique replacing fixed-weight residual connections in LLMs with learned attention-based aggregation, improving gradient distribution in the Kimi Linear 48B architecture.

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