Quanta AI Labs

12 / Mar 16 – 22, 2026 · 50 RELEASES

AI News Week 12, 2026

Week 12 2026 AI releases.

Language Models (4)

  • OpenViking

    ByteDance (VolcEngine) · Text Generation

    Open-source large language model project from ByteDance's cloud computing platform. Details unavailable due to page fetch failure.

  • MiniMax M2.7

    MiniMax · Reasoning

    Self-evolving agent model with native multi-agent collaboration. Scores 56.22% on SWE-Pro matching GPT-5.3-Codex, 66.6% medal rate on MLE Bench Lite, and 1495 ELO on GDPval-AA — highest among open-source models.

  • MiMo-V2-Pro

    Xiaomi · MoE Architecture

    Trillion-parameter foundation model (1T total, 42B active) optimized for agentic workloads with 1M-token context and hybrid attention. Ranks #2 among Chinese LLMs and #8 globally; achieves #3 on PinchBench (81.0) and coding performance approaching Claude Opus 4.6.

  • MiMo-V2-Omni

    Xiaomi · Reasoning

    Omni foundation model with unified image, video, and audio encoders designed for agentic systems. Demonstrates strong multimodal perception (surpassing Gemini 3 Pro on audio, approaching top closed-source models on vision) with native structured tool calling and real-time action capabilities for autonomous systems.

Vision & Image (3)

  • DreamID-Omni

    Guoxu1233 · Image-to-Image

    Multi-modal identity-preserving generation system enabling consistent character creation across images, videos, and 3D content with unified ID embeddings.

  • MiniCPM-SALA

    OpenBMB · OCR & Document

    Compact multimodal language model optimized for efficient inference with spatial-aware layout analysis capabilities for document understanding and visual reasoning tasks.

  • GlyphPrinter

    Fudan University, Nanyang Technological University · Text-to-Image

    A preference-based text rendering method using Region-Grouped Direct Preference Optimization (R-GDPO) to generate accurate glyphs in visual text synthesis, achieving fine-grained glyph accuracy for multilingual, complex, and out-of-domain characters through region-level preference annotations.

Video & Animation (5)

  • OmniLottie

    OpenVGLab · Text-to-Video

    First end-to-end multimodal Lottie animation generator (4B params). Creates vector animations from text, images, or video using VLMs. Accepted to CVPR 2026, trained on MMLottie-2M dataset.

  • SparkVSR

    Texas A&M University, YouTube/Google · Video Editing

    Interactive video super-resolution framework using sparse keyframe propagation with a Diffusion Transformer-based architecture. Achieves 24.6% improvement on CLIP-IQA and supports flexible keyframe-conditioned control for video restoration, old-film restoration, and style transfer tasks.

  • Dreamverse

    Hao AI Lab @ UCSD · Text-to-Video

    Real-time video generation interface enabling iterative "vibe directing" workflow, generating 5-second 1080p clips in ~4.55 seconds on single GPU (vs 1-2 minutes for Sora), powered by K2-V2 foundation model.

  • Seoul World Model

    KAIST AI, NAVER AI Lab, SNU AIIS · Text-to-Video

    A city-scale world model that generates photorealistic video of real Seoul streets over multi-kilometer trajectories using retrieval-augmented generation on 1.2M street-view images and synthetic CARLA data. Supports arbitrary camera navigation and text-prompted scenario control with a Virtual Lookahead Sink mechanism to prevent error accumulation.

  • ID-LoRA

    Tel Aviv University · Text-to-Video

    Generates identity-preserving video with synchronized audio from a reference image and audio clip. Built on LTX-2 diffusion model, preferred 73% over Kling 2.6 Pro and 81% over ElevenLabs for voice similarity.

Audio & Speech (3)

  • Granite 4.0 1B Speech

    IBM · Speech-to-Text

    A 1 billion parameter speech model from IBM's Granite family, designed for speech recognition and processing tasks with optimized inference for edge deployment.

  • Foundation-1

    RoyalCities · Music Generation

    Text-to-sample music generation model fine-tuned on Stable Audio Open 1.0. Generates production-ready loops with BPM, key, instrument, and timbral control across 40+ sonic descriptors.

  • RC Stable Audio Tools

    RoyalCities · Music Generation

    Tools and utilities for working with Stable Audio models, enabling audio generation and manipulation through accessible interfaces and scripts.

3D & Spatial (2)

  • SegviGen

    Beihang University, Renmin University of China, Tsinghua University, OriginArk · 3D Generation

    A 3D part segmentation framework that repurposes pretrained native 3D generative models through part-wise colorization on structured latent representations, achieving 40% improvement on interactive segmentation and 15% on full segmentation with 0.32% of typical training data.

  • SK-Adapter

    Hong Kong University of Science and Technology, University of Cambridge · 3D Generation

    A lightweight structural adapter network for skeleton-based control of native 3D generation that encodes joint coordinates and topology as learnable tokens injected via cross-attention into frozen 3D backbones. Enables precise structural articulation and local 3D editing with skeletal guidance while preserving geometry and texture quality.

Agents & Automation (14)

  • Skynet

    Ouro AI Labs · Agent Orchestration

    Open-source platform connecting multiple AI coding agents and humans into a shared real-time collaboration network. Supports Claude Code with adapter-based integration, WebSocket messaging, and isolated workspaces.

  • AutoResearch

    uditgoenka · Agent Framework

    GitHub repository for automated research workflows, likely enabling autonomous research task execution and literature analysis through AI agents.

  • ClawTeam

    HKUDS · Agent Orchestration

    Framework for AI agents to self-organize into collaborative swarms. Leader agents spawn workers in isolated git worktrees with task dependency management, inter-agent messaging, and adaptive resource allocation.

  • Agent Trenches Royale

    Agent Trenches · Agent Orchestration

    AI agent survival game where 10 agents with distinct personalities compete over 11 days on Solana blockchain. Agents trade tokens, launch narratives, and form alliances in a live game theory experiment.

  • IndusAGI

    varunisrani · Agent Framework

    All-in-one TypeScript SDK bundling AI model, agent, terminal UI, and web UI components. Includes a CLI coding agent and supports modular imports for selective integration.

  • IndusAGI Coding Agent

    varunisrani · Coding Agents

    An AI coding agent framework designed for automated code generation and software development tasks. Enables autonomous programming capabilities through agent-based architecture.

  • Mistral Vibe

    Mistral AI · Coding Agents

    Open-source CLI coding assistant powered by Mistral models. Features project-aware context, multiple agent profiles (plan, accept-edits, auto-approve), extensible skills system, and voice input support.

  • Clawless

    open-gitagent · Agent Framework

    Git agent framework enabling AI-driven repository automation and code management without direct file system access, operating through git commands and abstractions.

  • MedClaw

    MedClaw-Org · Agent Framework

    Personal AI medical assistant running Claude agents in isolated Linux containers. Includes PubMed search, biomedical database queries, USMLE prep, and connects via WhatsApp/Telegram with zero telemetry.

  • OpenMAIC

    THU-MAIC (Tsinghua University) · Agent Orchestration

    Multi-agent interaction and communication framework for collaborative AI systems. Provides infrastructure for coordinating and managing interactions between multiple AI agents.

  • MetaClaw

    Aiming Lab · Agent Framework

    Evolving AI agent framework that learns from live conversations via reinforcement learning. Acts as a transparent proxy injecting skills at each turn, with RL and MadMax modes for continuous fine-tuning.

  • LATENT

    Tsinghua University, Peking University, Galbot · Robotics

    A system for learning athletic humanoid tennis skills from imperfect human motion fragments using policy learning and sim-to-real transfer. Deployed on Unitree G1 robot, achieving multi-shot rallies with reactive footwork and natural motion styles.

  • MiroThinker 1.7-H1

    MiroMind · Agent Framework

    Heavy-duty research agent utilizing verification mechanisms to improve reasoning accuracy and reliability in complex research tasks. Focuses on enhanced verification-based approaches for autonomous research workflows.

  • Claude Code for Marketing

    Anthropic / HubSpot · Coding Agents

    Integration enabling Claude AI to assist with marketing tasks and code generation workflows through HubSpot's platform, leveraging Claude's code capabilities for marketing automation and content generation.

Developer Tools (13)

  • LLM Architecture Gallery

    Sebastian Raschka · AI Infrastructure

    Visual reference guide documenting transformer-based LLM architectures, showing layer configurations, attention mechanisms, and design patterns across different model families.

  • LookaheadKV

    Samsung Research · Inference Engine

    Lightweight KV cache eviction framework that glimpses into future tokens without expensive draft generation, reducing memory usage for long-context LLM inference while maintaining accuracy.

  • fastbook

    fastai · AI Infrastructure

    Jupyter notebooks and supplementary materials accompanying the fastai deep learning course, covering practical deep learning applications with PyTorch from foundational to advanced techniques.

  • Context Hub

    andrewyng · Prompt Engineering

    A framework for managing and organizing context in LLM applications, enabling efficient retrieval and injection of relevant information into prompts. Designed to improve context handling in RAG and multi-turn conversation systems.

  • CLIProxyAPI

    router-for-me · AI Infrastructure

    A CLI tool that provides proxy API functionality for routing and managing requests through command-line interfaces. Enables programmatic access to proxy services with configuration management.

  • Harvard Algorithmic Trading with AI

    moondevonyt · AI Infrastructure

    Educational repository implementing algorithmic trading strategies using AI/ML techniques, likely covering model development, backtesting, and quantitative finance applications.

  • OpenShell

    NVIDIA · AI Infrastructure

    Sandboxed execution environment for autonomous AI agents with declarative YAML policies. Enforces filesystem, network, process, and inference routing restrictions. Supports Claude Code, Codex, and Copilot CLI.

  • Workforce Data Initiative

    Workforce Data Initiative · Data Integration

    Open data infrastructure for workforce development featuring skills APIs, ML pipelines for job posting analysis, and training program registries. Powers dataatwork.org.

  • Unsloth

    UnslothAI · AI Infrastructure

    Framework for efficient fine-tuning of large language models with 2-5x speedup and 80% reduced memory usage. Optimizes training through kernel fusion and algorithmic improvements for consumer hardware.

  • ComfyUI-Foundation-1

    Saganaki22 · AI Infrastructure

    ComfyUI extension or node pack for Foundation model integration, enabling visual node-based workflows for foundation model inference and image generation tasks.

  • ComfyUI-Forbidden-Vision

    luxdelux7 · AI Infrastructure

    ComfyUI extension for advanced vision-based image processing and manipulation workflows, integrating custom vision models with the ComfyUI node-based interface for generative tasks.

  • Alireza Rezvani

    Alireza Rezvani · Prompt Engineering

    CTO and AI engineer behind claude-skills (6.2k stars, 192+ Claude Code skills), claude-code-skill-factory, and claude-code-tresor. Prolific contributor to the Claude Code agent ecosystem.

  • Google AI Studio

    Google · AI Infrastructure

    Web-based IDE for interacting with Google's generative AI models, enabling prompt engineering and API integration for developers.

Research & Safety (6)

  • OBLITERATUS

    elder-plinius · Red Teaming

    Toolkit for understanding and removing refusal behaviors from LLMs via abliteration. Uses SVD-based extraction of refusal directions with a 6-stage pipeline and 15 analysis modules.

  • hooeem AI Updates

    hooeem · Survey & Overview

    AI community updates and discussions on latest developments in the AI ecosystem, shared via X/Twitter.

  • CS146S: The Modern Software Developer

    Stanford University · Survey & Overview

    A Stanford computer science course covering modern software development practices and techniques relevant to contemporary software engineering. Appears to be an educational curriculum rather than a specific AI product or release.

  • Evo2

    Arc Institute · Research Paper

    DNA foundation model trained on 8.8 trillion tokens from all domains of life. Models sequences at single-nucleotide resolution up to 1M base pair context. Available in 1B to 40B parameter sizes.

  • Four Arms One Monster

    OpenSourceMalware.com · Red Teaming

    Community threat intelligence analysis documenting a multi-component malware variant with four distinct attack vectors or propagation mechanisms. Technical breakdown of detection signatures and behavioral indicators.

  • Terminator

    UT Austin, EPFL, ENS Paris-Saclay, Télécom Paris · Research Paper

    Early-exit strategy for Large Reasoning Models that reduces Chain-of-Thought reasoning length by 14-55% and halves wall-clock latency by training a lightweight single-layer probe to predict when the model has generated its final answer, achieving best or second-best performance on 28/32 metrics across four models and benchmarks without fine-tuning.

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