Quanta AI Labs

09 / Feb 23 – Mar 1, 2026 · 56 RELEASES

AI News Week 9, 2026

Qwen3.5 expands with 122B-A10B, 35B-A3B, and 27B variants. Agent Skills emerge as a cross-platform standard backed by Anthropic, OpenAI, and HuggingFace. Anthropic exposes industrial-scale distillation attacks. Robotics surges with AGIBOT A2, Unitree R1, and NVIDIA EgoScale. TTS innovation continues with Voicebox, Faster Qwen3-TTS, and NeuTTS Nano.

Language Models (5)

  • Qwen3.5-122B-A10B

    Alibaba Qwen · MoE Architecture

    Multimodal MoE model with 122B total / 10B active params. Hybrid Gated DeltaNet + sparse MoE architecture with 256 experts. Native 262K context (extensible to 1M+) with vision, image, and video understanding.

  • Qwen3.5-35B-A3B

    Alibaba Qwen · MoE Architecture

    Multimodal MoE with 35B total / 3B active params. 40-layer hybrid architecture combining Gated DeltaNet and sparse MoE (256 experts). Supports vision-language tasks with 262K native context.

  • Qwen3.5-27B

    Alibaba Qwen · Text Generation

    27B dense multimodal model with unified vision-language. 64-layer hybrid Gated DeltaNet + FFN architecture with 262K native context, extensible to 1M+ via YaRN scaling. Multi-Token Prediction training.

  • LFM2-24B-A2B

    Liquid AI · MoE Architecture

    Hybrid MoE model with 24B total / 2.3B active params designed for efficient on-device deployment. General-purpose conversational AI with tool use capabilities.

  • Mercury Chat

    Inception Labs · Text Generation

    Chat interface powered by Mercury, a diffusion-based LLM that generates text in parallel (non-autoregressive) rather than token-by-token for significantly faster inference.

Vision & Image (6)

  • GLM-OCR

    ZhipuAI · OCR & Document

    Multimodal OCR model for complex document understanding achieving state-of-the-art performance on document parsing and information extraction tasks.

  • Nemotron ColEmbed VL 4B v2

    NVIDIA · Visual Understanding

    State-of-the-art late interaction embedding model for visual document retrieval. Takes text queries and document images as input, outputs multi-vector representations for retrieval tasks.

  • QuiverAI

    QuiverAI · Text-to-Image

    AI-powered platform for generating, editing, and animating editable vector graphics from text prompts and images. Backed by $8.3M a16z seed funding.

  • VecGlypher

    Meta AI / UC Santa Cruz · Text-to-Image

    Multimodal language model that generates editable SVG font glyphs directly from text prompts or reference images without raster intermediates.

  • LoRWeB

    NVIDIA / Technion · Image Editing

    Modular framework that dynamically combines learned low-rank adapters to perform visual image transformations from example pairs, enabling flexible edits without test-time optimization.

  • PhysicEdit

    KAUST / CUHK / Hugging Face · Image Editing

    Image editing framework treating edits as physical state transitions rather than discrete mappings, enabling realistic rendering of refraction and material deformation.

Video & Animation (3)

  • DreamID-Omni

    Tsinghua University · Video Editing

    Unified framework for generation, editing, and animation of synchronized human audio-video content using reference images and voice timbres.

  • Solaris

    NYU VisionX · Text-to-Video

    First multiplayer video world model generating consistent observations across multiple players simultaneously in Minecraft.

  • VidEoMT

    TU Eindhoven / RWTH Aachen · Video Understanding

    Encoder-only video segmentation model performing temporal tracking through lightweight query propagation. 5–10x faster than traditional approaches with competitive accuracy.

Audio & Speech (6)

  • Voicebox

    Jamie Pine · Voice Cloning

    Local-first open-source voice cloning and synthesis studio. Built with Tauri (Rust), React, and Qwen3-TTS. Features multi-track timeline editing, voice profiles, and REST API. Alternative to ElevenLabs.

  • Faster Qwen3-TTS

    andimarafioti · Text-to-Speech

    Optimized TTS inference using CUDA graph capture for Qwen3-TTS. Up to 9.8x speedup on RTX 4060 without Flash Attention or vLLM. Supports streaming audio and voice cloning (0.6B and 1.7B models).

  • NeuTTS Nano Multilingual

    Neuphonic · Text-to-Speech

    Lightweight TTS model 3x smaller than NeuTTS Air. Runs in real-time on CPU with support for English, Spanish, French, and German.

  • Sopro TTS v1.5

    Samuel Vitorino · Text-to-Speech

    Lightweight 135M parameter English TTS model with streaming and zero-shot voice cloning. Achieves 0.05 real-time factor on CPU.

  • LavaSR

    Yatharth Sharma · Audio Processing

    Lightweight speech enhancement model that restores degraded audio to high quality at extreme speed — 5000x realtime on GPUs.

  • Echoes Over Time

    Sony AI · Music Generation

    MMHNet framework that enables video-to-audio generation models to produce long-form audio (5+ minutes) by training on shorter sequences. Unlocks length generalization.

3D & Spatial (2)

  • Generated Reality

    Stanford / NYU Shanghai · 3D Generation

    Video world model generating interactive egocentric virtual environments in real-time by tracking user hand and head poses through VR hardware.

  • tttLRM

    UPenn / Adobe Research · 3D Reconstruction

    Test-Time Training for 3D reconstruction enabling efficient autoregressive 3D object and scene reconstruction from multiple image observations.

Agents & Automation (16)

  • LocoOperator-4B

    LocoreMind · Coding Agents

    4B-parameter tool-calling agent model distilled from Qwen3-Coder-Next. Specializes in multi-turn codebase exploration with perfect JSON validity for structured tool calls.

  • SARAH

    Meta Reality Labs · Embodied AI

    Real-time method generating full-body conversational motion for embodied agents that responds to user position and speech with natural spatial awareness and gaze behavior.

  • EgoScale

    NVIDIA GEAR Lab · Robotics

    Framework training robotic hands for complex manipulation using 20,000+ hours of human egocentric video combined with vision-language-action models for human-to-robot transfer.

  • AGIBOT A2

    AGIBOT · Robotics

    169cm humanoid service robot with 40+ degrees of freedom, dexterous hands, and L4-level autonomous mobility for reception, marketing, and presentation roles.

  • Unitree R1

    Unitree Robotics · Robotics

    Consumer humanoid robot from $4,900 joining Unitree's lineup alongside Go2 quadrupeds ($1,600) and industrial H1 humanoids ($90,000).

  • OpenClaw-RL

    Gen-Verse · Agent Framework

    Fully async reinforcement learning framework personalizing AI agents through natural conversation feedback. Supports Binary RL and On-Policy Distillation without external API keys.

  • memclawz

    Yoni Assia · Agent Framework

    Three-layer hybrid memory system for OpenClaw agents: fast working memory (QMD), vector search (Zvec), and automatic compaction. Runs locally without external APIs.

  • MimiClaw

    memovai · Embodied AI

    Pocket-sized AI assistant on a $5 ESP32-S3 chip (0.5W). Interacts with Claude/GPT via Telegram with persistent local flash memory, cron scheduling, and autonomous task execution.

  • OpenFang

    RightNow AI · Agent Orchestration

    Open-source Agent Operating System built in Rust with 7 autonomous Hands running on schedules, building knowledge graphs, and reporting to a dashboard.

  • Ruflo

    ruvnet · Agent Orchestration

    Enterprise AI orchestration platform deploying 60+ specialized agents in coordinated swarms with self-learning, fault-tolerant consensus, and native Claude Code integration.

  • Dev Agent (Build-app-like-pro)

    First-Principal Labs · Coding Agents

    Autonomous Python agent transforming project ideas into fully-implemented apps by automating planning, GitHub issue creation, feature development with Claude Code CLI, and PR management.

  • Cheatcode

    cheatcode-ai · Coding Agents

    Open-source AI coding agent for building, previewing, and deploying complete web and mobile apps. Next.js + FastAPI with multi-LLM support (OpenAI, Anthropic, OpenRouter).

  • Claude Financial Services Plugins

    Anthropic · Agent Framework

    41 skills, 38 commands, and 11 MCP integrations for investment banking, equity research, private equity, and wealth management. Builds DCF, LBO, and 3-statement models with live Excel formulas.

  • ConnectWithAllCode

    IAmUnbounded · Coding Agents

    Locally-run WhatsApp interface connecting multiple coding agents (Claude Code, Gemini, Cursor, Windsurf) through a gateway and connector daemon.

  • Agent Skills Standard

    Anthropic / agentskills.io · Agent Framework

    Open format for giving agents reusable capabilities. Supported by Claude Code, Cursor, Gemini CLI, OpenAI Codex, GitHub, VS Code, and 20+ agent platforms.

  • Anthropic Skills

    Anthropic · Agent Framework

    Collection of 16 reusable Claude skills including PDF generation, PPTX/XLSX export, MCP builder, frontend design, webapp testing, and skill creator.

Developer Tools (13)

  • OpenAI Skills

    OpenAI · AI Infrastructure

    Reusable skill definitions and templates for OpenAI's Codex. Organized into .system (auto-install), .curated (named install), and .experimental categories.

  • Hugging Face Skills

    Hugging Face · AI Infrastructure

    Standardized AI/ML task definitions working across Claude Code, Codex, Gemini CLI, and Cursor. Covers dataset creation, model training, evaluation, and paper publishing.

  • The Agent Skills Directory

    Vercel Labs · AI Infrastructure

    Community hub for discovering and installing reusable AI agent skills with a single command. Searchable directory of capabilities for coding agents.

  • Trigger.dev

    Trigger.dev · AI Infrastructure

    Platform for building and deploying fully-managed AI agents and long-running TypeScript workflows with built-in retries, queuing, observability, and elastic scaling.

  • Modal

    Modal Labs · AI Infrastructure

    Cloud infrastructure for deploying and scaling AI workloads — inference, training, and batch processing — using Python with minimal configuration.

  • Modal Sandbox Agent

    Modal Labs · AI Infrastructure

    Example showing how to run Claude Code inside isolated Modal Sandbox environments for safe repository analysis and code execution.

  • Opik

    Comet ML · AI Infrastructure

    Open-source platform for debugging, evaluating, and monitoring LLM applications with comprehensive tracing and automated evaluations.

  • W&B Serverless SFT

    Weights & Biases · AI Infrastructure

    Managed fine-tuning service training low-rank adapters on curated datasets with automatic infrastructure provisioning and model deployment.

  • Topic to Visual

    First-Principal Labs · Prompt Engineering

    AI-powered web app converting technical concepts into interactive learning experiences with explanations, use cases, and sandboxed playgrounds. Built with React 19 and GPT.

  • Ultralytics Platform

    Ultralytics · AI Infrastructure

    Community hub for discovering and sharing YOLO computer vision models and datasets with cloud-based training and deployment.

  • Free LLM API Resources

    cheahjs · AI Infrastructure

    Curated list of services providing free access or trial credits for API-based LLM usage, including OpenRouter, Google AI Studio, Fireworks, and more.

  • NVIDIA Generative AI Examples

    NVIDIA · AI Infrastructure

    Reference workflows and examples for building generative AI systems using NVIDIA's ecosystem, including RAG pipelines, agentic workflows, and model fine-tuning.

  • OpenAlgo Indicator Skills

    marketcalls · BI & Analytics

    100+ Numba-optimized technical indicators with Plotly charting, Dash/Streamlit dashboards, and real-time WebSocket support for Indian and US markets.

Research & Safety (5)

  • Detecting and Preventing Distillation Attacks

    Anthropic · Alignment & Safety

    Anthropic discloses industrial-scale distillation campaigns by DeepSeek, Moonshot, and MiniMax involving over 16 million fraudulent API exchanges to extract Claude's capabilities.

  • Doc-to-LoRA

    Sakana AI · Research Paper

    Hypernetwork-based method that generates compact LoRA adapters in a single forward pass for instant knowledge internalization, replacing expensive per-task optimization with sub-second forward passes.

  • UnifiedReward-Flex

    Fudan / Shanghai AI Lab · Research Paper

    Unified reward model enabling personalized reasoning for improving both image and video generation through preference-based optimization.

  • VBVR

    Video-Reason · Benchmark

    Comprehensive benchmark and dataset for evaluating video generation models on spatiotemporal reasoning tasks across multiple cognitive categories.

  • MicroEDR

    abshkd · Alignment & Safety

    Lightweight endpoint detection and response system for Windows servers using ETW tracing. Automated ransomware detection heuristic triggers process termination on suspicious file-write patterns.

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