Quanta AI Labs

11 / Mar 9 – Mar 15, 2026 · 41 RELEASES

AI News Week 11, 2026

NVIDIA Nemotron-3-Super 120B leads frontier with hybrid Mamba-MoE and 1M context. Fish Audio S2 and Hume TADA open-source TTS. Hermes Agent, ClawWork, and Vercel Agent Skills surge in agentic tooling. ComfyUI launches App Mode and ComfyHub. FLUX.2-klein adds KV-cache for 2.5x faster editing.

Language Models (2)

  • Nemotron-3-Super-120B-A12B

    NVIDIA · MoE Architecture

    Hybrid Mamba-2 + MoE + Attention model with 120B total / 12B active params and 1M-token context. Multi-Token Prediction, 3-stage RLHF training, NVFP4 quantization. 90.21 AIME25, 60.47 SWE-Bench, 91.75 RULER@1M. Supports 7 languages.

  • Tri-21B-Think

    Trillion Labs · Reasoning

    20.7B-param chain-of-thought reasoning model with 262K context (YaRN scaling). SFT + RL reasoning pipeline optimized for English, Korean, and Japanese. 62.6 GPQA-Diamond, 56.67 AIME 2026, 81 TAU2-Bench.

Vision & Image (3)

  • Penguin-VL-2B

    Tencent · Visual Understanding

    Compact 2B VLM that initializes its vision encoder from Qwen3-0.6B (not CLIP/SigLIP) to reduce objective mismatch. Temporal Redundancy-Aware (TRA) token compression for video. Outperforms Qwen3-VL-2B on DocVQA (94.1), ChartQA (86.6), and LongVideoBench (59.5).

  • FLUX.2-klein-9B-KV

    Black Forest Labs · Text-to-Image

    9B image generation and editing model with step-distilled 4-step inference and KV-cache that reduces multi-reference editing compute by 2.5x. Supports text-to-image, image-to-image, and multi-reference editing on ~29GB VRAM.

  • Anima

    CircleStone Labs / Comfy Org · Text-to-Image

    2B anime and illustration generation model built on NVIDIA Cosmos-Predict2 with Qwen3 text encoder. Trained on several million anime images and 800K artistic images, no synthetic data. Supports Danbooru-style tags, artist tags, and natural language.

Video & Animation (7)

  • MatAnyone 2

    NTU S-Lab / SenseTime · Video Editing

    CVPR 2026 video matting framework with a learned Matting Quality Evaluator (MQE) for online training feedback without ground truth. Introduces VMReal dataset with 28K clips and 2.4M frames. Reference-frame training for handling appearance variations in long videos.

  • WildActor

    Research · Text-to-Video

    Human video generation framework preserving consistent full-body identity across varying camera angles and motions. Introduces Actor-18M dataset (1.6M videos, 18M images), Asymmetric Identity-Preserving Attention (AIPA), and Viewpoint-Adaptive Monte Carlo Sampling.

  • ShotVerse

    Research · Text-to-Video

    Text-driven multi-shot cinematic video generation with precise camera control. Plan-then-Control VLM architecture, automated camera calibration pipeline, and 4D RoPE capturing shot hierarchy. Introduces ShotVerse-Bench for evaluation.

  • EffectMaker

    Research · Video Editing

    Zero-shot visual effects generation from reference VFX videos without per-effect fine-tuning. Dual-Path Guidance combining MLLM semantic understanding with DiT in-context learning. Introduces EffectData: 120K videos across 3K VFX categories.

  • OmniLottie

    OpenVGLab (CVPR 2026) · Video Editing

    CVPR 2026 multimodal instructed vector animation generator producing Lottie JSON files. Open-source VLM-based system for generating scalable, resolution-independent animated vectors from text or image prompts.

  • Diagonal Distillation

    SphereLab AI · Text-to-Video

    Video diffusion distillation achieving 277x speedup — generates 5-second videos in 2.61 seconds. Asymmetric denoising steps, exposure bias correction, and latent-space optical flow for chunk-based generation without per-step VAE reconstruction.

  • BrandFusion

    Research · Text-to-Video

    Automatic brand embedding into AI-generated videos for sustainable T2V monetization. Offline brand knowledge base construction + online 5-agent pipeline (Selector, Strategy Generator, Prompt Refiner, Critic, Experience Learner) for seamless brand integration.

Audio & Speech (3)

  • MagpieTTS Multilingual 357M

    NVIDIA · Text-to-Speech

    357M-param multilingual TTS model supporting 9 languages (English, Spanish, German, French, Vietnamese, Italian, Mandarin, Hindi, Japanese) across 5 voice personas. Causal encoder-decoder with CFG and GRPO training. Up to 20s generation, commercial-ready.

  • TADA (Open Source)

    Hume AI · Text-to-Speech

    Open-source 1B and 3B Llama-based TTS with 1:1 text-audio token alignment eliminating hallucinations by design. 0.09 RTF (5x faster than comparable systems), 4.18/5.0 speaker similarity, ~700s audio per standard context budget. Apache 2.0.

  • Fish Audio S2 (Open Source)

    Fish Audio · Text-to-Speech

    Open-source TTS model trained on 10M+ hours across 50 languages. Dual-AR architecture (slow semantic + fast acoustic), inline prosody/emotion tags ([laugh], [whispers]), 0.195 RTF on H200, ~100ms TTFA, 3,000+ tokens/sec. Includes fine-tuning code and SGLang streaming engine.

3D & Spatial (6)

  • InSpatio WorldFM

    InSpatio · 3D Generation

    Real-time generative frame model for interactive 3D world exploration from single images. Hybrid memory (explicit 3D anchors + implicit neural memory) for multi-view consistency. Spatial editing via text, runs on consumer GPUs.

  • RL3DEdit

    AMAP-ML / Alibaba · 3D Generation

    Reinforcement learning approach to instruction-based 3D scene editing. Trains editing policies through environment feedback to achieve controllable, consistent 3D scene modifications from natural language instructions.

  • Mobile-GS

    Research · NeRF & Gaussian

    Real-time Gaussian Splatting on mobile devices at 116+ FPS on Snapdragon 8 Gen 3. Depth-aware order-independent rendering (eliminates costly sorting), neural view-dependent MLP, neural vector quantization, and contribution-based pruning to 4.8MB model size.

  • Holi-Spatial

    Research · 3D Reconstruction

    Egocentric video-to-3D spatial understanding framework processing continuous video streams into coherent spatial trajectories. 13K+ scenes, 1.3M+ QA pairs, 300K+ 3D bounding boxes benchmark. Grounding + depth coupling for spatial QA.

  • LoGeR

    Research · 3D Reconstruction

    Long-context 3D geometric reconstruction of up to 19,000-frame video sequences without post-processing optimization. Hybrid memory: Sliding Window Attention (SWA) for local alignment + Test-Time Training (TTT) for global consistency at sub-quadratic scaling.

  • Orient Anything V2

    SpatialVision (NeurIPS 2025 Spotlight) · Depth Estimation

    NeurIPS 2025 Spotlight model for 3D object orientation estimation from a single image. Predicts full 6-DOF pose including viewpoint-invariant canonical orientation for any object category.

Agents & Automation (9)

  • OmniCoder-9B

    Tesslate · Coding Agents

    Coding agent fine-tuned on 425K+ agentic trajectories based on Qwen3.5-9B with 262K context. Read-before-write patterns, LSP diagnostics, minimal edit diffs. 83.8% GPQA Diamond, 23.6% Terminal-Bench 2.0 (+61% over base). Apache 2.0.

  • HKUDS Lab Projects (LightRAG, CLI-Anything, ClawWork)

    HKUDS / HKU · Agent Framework

    Data Intelligence Lab at HKU releases three major open-source projects: LightRAG (29K+ stars), CLI-Anything (making all software agent-native, 14.7K+ stars), and ClawWork (OpenClaw AI coworker, $15K earned in 11 hours, 7.2K+ stars).

  • Atropos

    Nous Research · Agent Framework

    LLM Reinforcement Learning Environments framework for collecting and evaluating LLM trajectories across diverse environments. Enables multi-environment RLHF with asynchronous rollout collection. 899 GitHub stars.

  • Vercel Agent Browser

    Vercel Labs · Web Agents

    Browser automation CLI for AI agents enabling programmatic web browsing, scraping, and interaction. 22K+ GitHub stars. Pairs with Vercel Agent Skills collection (23K+ stars) for composable agent capabilities.

  • ClawWork

    HKUDS · Agent Framework

    Framework turning OpenClaw into an AI coworker that autonomously handles professional tasks and earns real revenue. Demo: $15K earned in 11 hours. 7.2K+ GitHub stars.

  • CLI-Anything

    HKUDS · Agent Framework

    Makes ALL software agent-native by exposing any CLI tool through a unified agent interface. Enables AI agents to use arbitrary command-line software without custom integrations. 14.7K+ GitHub stars.

  • Gemini Robotics-ER 1.5

    Google · Robotics

    VLM for physical robots with object detection, trajectory planning, and multi-step task orchestration via natural language. Processes images, video, and audio. Generates function calls for custom robot APIs with 1M-token context.

  • Claude Scientific Skills

    K-Dense AI · Agent Framework

    Ready-to-use agent skills for research, science, engineering, analysis, finance, and writing covering bioinformatics, drug discovery, genomics, computational biology, and materials science. 15K+ GitHub stars.

  • pi-mono

    badlogic · Agent Framework

    AI agent toolkit monorepo: coding agent CLI, unified LLM API, TUI and web UI libraries, Slack bot, and vLLM pod management. Single-command deployments for full agent stacks. 24K+ GitHub stars.

Developer Tools (8)

  • Gemini Embedding 2

    Google · AI Infrastructure

    Google's first natively multimodal embedding model mapping text, images, video, audio, and documents into a single unified vector space — enabling cross-modal retrieval and similarity across all content types.

  • Claude Code Action

    Anthropic · AI Infrastructure

    Official GitHub Action running Claude Code directly in CI/CD workflows. Automates code reviews, issue triaging, PR generation, and test fixing triggered by GitHub events. 6.3K+ GitHub stars.

  • AngelSlim

    Tencent · AI Infrastructure

    Comprehensive model compression toolkit for LLMs and VLMs supporting quantization (FP4, MXFP4), speculative decoding (EAGLE), and pruning. Covers HunyuanVideo, DeepSeek, and Qwen models with enhanced usability and efficiency.

  • PageIndex

    VectifyAI · Data Integration

    Vectorless, reasoning-based RAG using document page indexes instead of embeddings. LLM reads the index to navigate to relevant pages, then retrieves exact content. 21.5K+ GitHub stars.

  • Plano

    Katanemo · AI Infrastructure

    AI-native proxy and data plane for agentic apps with built-in orchestration, safety, observability, and smart LLM routing via Envoy proxy. Keeps agent logic clean by handling cross-cutting concerns at the infrastructure layer. 6K+ GitHub stars.

  • Google Workspace CLI

    Google Workspace · AI Infrastructure

    One command-line tool for Drive, Gmail, Calendar, Sheets, Docs, Chat, Admin, and more. Dynamically built from Google Discovery Service in Rust. Includes AI agent skills and Gemini CLI extension. 20.5K+ GitHub stars.

  • Claude Replay

    es617 · AI Infrastructure

    Converts AI coding agent sessions (Claude Code, Cursor, Codex) into self-contained, embeddable HTML replays for sharing, debugging, and team review. 510 GitHub stars.

  • ComfyUI App Mode & ComfyHub

    Comfy Org · AI Infrastructure

    ComfyUI launches App Mode (node graph → clean UI), App Builder (configure inputs/outputs), shareable app URLs, and ComfyHub marketplace for community workflow and app sharing. Available on Comfy Cloud and local.

Research & Safety (3)

  • AGENTS.md Evaluation Study

    ETH Zurich / DeepMind · Research Paper

    First rigorous empirical study showing repository-level context files like AGENTS.md reduce coding agent task success rates vs. no context, while increasing costs by 20%+. Recommends minimal, constraint-focused documentation for agent workflows.

  • LangChain: Coding Agents Report

    LangChain · Survey & Overview

    Analysis showing coding agents shift the bottleneck from implementation to review, split teams into builders and reviewers, and transform PRDs into structured prompts. Argues diverse backgrounds benefit equally if they develop product sense.

  • Transformer Explainer

    Georgia Tech (Polo Club) · Interpretability

    Interactive browser-based tool running GPT-2 via ONNX Runtime to visualize transformer mechanics: attention maps, embeddings, and token predictions. Temperature, top-k, and top-p controls. Built with Svelte and D3.js.

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