Quanta AI Labs

10 / Mar 2 – Mar 8, 2026 · 53 RELEASES

AI News Week 10, 2026

GPT-5.4 launches with 1M-token context and native computer-use. Sarvam-105B open-sources a 106B MoE for 22 Indian languages. Immersive media surges with Helios real-time video gen, Kiwi-Edit, and LTX-2.3. CUDA Agent achieves 2.11x GPU kernel speedup via agentic RL. WorldMonitor and GitNexus lead developer tooling.

Language Models (3)

  • GPT-5.4

    OpenAI · Text Generation

    OpenAI's most capable unified model with native computer-use, 1M-token context window, and 128K max output. Available in Standard, Thinking, and Pro variants. 33% fewer false claims per response vs GPT-5.2.

  • Qwen3.5 Collection

    Alibaba Qwen · Multilingual

    Full Qwen3.5 model family with 24 variants from 0.8B to 397B-A17B. All are multimodal image-text-to-text models with hybrid Gated DeltaNet + sparse MoE architecture and 262K native context.

  • Sarvam-105B

    Sarvam AI · Multilingual

    Open-source MoE model with 106B total / 10.3B active params supporting 22 Indian languages. Top-8 routing over 128 experts with 128K context. Scores 90.6 MMLU, 98.6 Math500, and 96.7 on AIME 25 with tools.

Vision & Image (5)

  • Phi-4-Reasoning-Vision-15B

    Microsoft · Visual Understanding

    15B multimodal reasoning model with SigLIP-2 vision encoder and mid-fusion architecture. Dynamic THINK/NOTHINK modes for selective reasoning. 88.2% on ScreenSpot-V2 GUI grounding, trained on 240 B200s.

  • FireRed-Image-Edit-1.1

    FireRedTeam · Image Editing

    Open-source image editing foundation model with SOTA identity consistency, multi-element fusion (10+ elements), portrait makeup, and photo restoration. 4.5s end-to-end generation with 30GB VRAM. Apache 2.0 license.

  • HY-WU (Hunyuan Functional Memory)

    Tencent Hunyuan · Image Editing

    Generates instance-conditioned LoRA adapters on-the-fly without fine-tuning for personalized image editing. Handles clothing fusion, character outfit migration, face identity transfer, and virtual try-on at inference time.

  • SpatialT2I

    PKU DA Group · Text-to-Image

    Enhances spatial understanding in text-to-image models using SpatialScore reward model trained on 80K+ preference pairs. GRPO online RL pipeline improves spatial accuracy over Flux.1-dev baselines.

  • HiFi-Inpaint

    Research · Image Editing

    DiT-based framework for seamlessly compositing product images into human photos for advertising and e-commerce. Shared Enhancement Attention (SEA) and Detail-Aware Loss preserve fine textures and text. Includes HP-Image-40K dataset.

Video & Animation (7)

  • LTX-2.3 Video Engine

    LTX · Text-to-Video

    DiT-based video generation model with synchronized audio from text, image, or audio inputs. Up to 1080p native portrait video, 20-second generation, and redesigned VAE for sharper detail. Open-source on HuggingFace.

  • Kiwi-Edit

    ShowLab / NUS · Video Editing

    Video editing framework using natural language instructions and reference images at 720p. Combines Multimodal LLM with video diffusion transformer. 3-stage training curriculum on 477K reference-guided quadruplets.

  • Helios

    PKU Yuan Group · Text-to-Video

    14B autoregressive diffusion model generating minute-scale videos at 19.5 FPS on a single H100. Addresses video drift without self-forcing or keyframe sampling through drift-simulating training strategies.

  • FREE-Edit

    Research · Video Editing

    Zero-shot video editing that propagates first-frame edits throughout a video without fine-tuning. Editing-awaRE (REE) injection adaptively modulates token intensity using optical flow to track edited regions.

  • RealWonder

    Research · Text-to-Video

    Real-time physics-conditioned video generation at 13.2 FPS from a single image. Uses physics simulation as intermediary to translate forces and robotic manipulations into visual outputs across rigid, deformable, fluid, and granular materials.

  • Varg SDK

    VargHQ · Video Editing

    AI-native SDK for video tooling enabling programmatic video processing, editing, and generation workflows through a developer-friendly API.

  • CubeComposer

    Research · Text-to-Video

    Spatio-temporal autoregressive diffusion model that natively generates 4K 360-degree videos. Decomposes video into cubemap faces with sparse context attention, cube-aware positional encoding, and continuity blending to eliminate boundary seams.

Audio & Speech (1)

  • Sarvam Voice Agent with Pipecat

    Sarvam AI · Voice Agents

    Integration guide for building voice AI agents using Sarvam's TTS/STT APIs with the Pipecat framework. Supports 22 Indian languages with real-time voice interaction capabilities.

3D & Spatial (5)

  • DiffusionHarmonizer

    NVIDIA · NeRF & Gaussian

    Single-step temporally-conditioned enhancer that converts NeRF/3DGS renderings into temporally consistent, realistic outputs for robot simulation and autonomous vehicle testing. Runs on a single GPU during online simulation.

  • ArtiFixer

    Research · 3D Reconstruction

    Enhances and extends 3D scene reconstruction by generating plausible novel views in unobserved areas. Combines autoregressive diffusion with 3D representations for scene generation from sparse reconstructions or text prompts.

  • Track4World

    Research · 3D Reconstruction

    Feedforward world-centric dense 3D tracking of all pixels from monocular video. Novel 3D correlation scheme simultaneously estimates pixel-wise 2D and 3D dense flow between arbitrary frame pairs using a VGGT-style ViT.

  • Utonia

    HKU / Xiaomi · Point Cloud

    Unified self-supervised encoder for all point cloud domains (remote sensing, LiDAR, indoor, CAD, video). Introduces Causal Modality Blinding, Perceptual Granularity Rescale, and RoPE for cross-domain spatial encoding.

  • Git City

    srizzon · 3D Generation

    Visualize GitHub profiles as 3D pixel art buildings in an interactive city. Built with Next.js, React Three Fiber, Three.js, and Supabase. 3K+ GitHub stars.

Agents & Automation (11)

  • UI-Venus-1.5-2B

    InclusionAI · Web Agents

    Unified 2B GUI agent for web, mobile, and desktop built on Qwen3-VL. 4-stage training pipeline (Mid-Training, Offline-RL, Online-RL, Model Merging) across 10B tokens and 30+ GUI datasets. 57.7% on ScreenSpot-Pro.

  • OmniXtreme

    BIGAI / Unitree · Robotics

    Unified humanoid robot policy for extreme acrobatic movements including backflips, handstands, and combat sequences. DAgger-based Flow Matching with Power-Safety Regularization prevents unsafe energy absorption.

  • CUDA Agent

    ByteDance Seed / Tsinghua · Agent Framework

    Large-scale agentic RL system that optimizes GPU kernels, achieving 2.11x speedup over torch.compile on KernelBench. Staged RL training (warmup, rejection fine-tuning, multi-turn optimization) up to 128K context.

  • Augment Code

    Augment · Coding Agents

    AI coding platform with a Context Engine that maintains live understanding of entire codebases including dependencies, architecture, and history. IDE agents (VS Code, JetBrains), CLI tool, and automated code review.

  • GitAgent

    GitAgent · Agent Framework

    Open standard for defining and managing AI agents as version-controlled files in git repos. Framework-agnostic (Claude, OpenAI, CrewAI), MIT-licensed CLI for scaffolding, validation, and export to multiple runtimes.

  • TEN Framework

    TEN Framework · Agent Framework

    Open-source framework for building conversational voice AI agents with real-time multi-modal capabilities. Supports AI-powered voice, video, and interactive applications. 10K+ GitHub stars.

  • memU

    NevaMind AI · Agent Orchestration

    Memory system for 24/7 proactive agents like OpenClaw with MCP integration, sandbox capabilities, and Claude Skills support. Designed for persistent, context-aware agent workflows. 12K+ GitHub stars.

  • Memento

    Agent-on-the-Fly · Agent Framework

    Fine-tunes LLM agents without fine-tuning the underlying LLMs. Enables efficient agent behavior specialization while preserving the base model's general capabilities. 2.3K+ GitHub stars.

  • Claude Code Best Practice

    shanraisshan · Agent Framework

    Comprehensive guide to Claude Code agentic engineering best practices, covering prompting patterns, workflows, and optimization techniques. 12K+ GitHub stars.

  • Pixel Agents

    pablodelucca · Agent Framework

    Pixel-art office simulation featuring autonomous AI agents in an interactive pixel environment. Visual sandbox for agent behavior experimentation. 3.6K+ GitHub stars.

  • PinchTab

    pinchtab · Web Agents

    High-performance browser automation bridge and multi-instance orchestrator with advanced stealth injection and real-time dashboard. Built in Go with CDP support. 5.5K+ GitHub stars.

Developer Tools (17)

  • Jina Embeddings V5 Text-Small

    Jina AI · AI Infrastructure

    677M-param embedding model based on Qwen3-0.6B-Base with 1024-dim Matryoshka embeddings. Supports 119+ languages and 32K tokens. Highest performing multilingual embedding model under 1B parameters.

  • Spectrum

    Research · Inference Engine

    Accelerates diffusion model inference up to 4.79x on FLUX.1 by forecasting features using Chebyshev polynomials in the spectral domain. Global long-range feature reuse with non-compounding error bounds.

  • Turbopuffer

    Turbopuffer · AI Infrastructure

    Serverless vector and full-text search database on object storage with sub-10ms p50 latency. Handles 2.5T+ documents, 10M+ writes/sec, 10K+ queries/sec in production. Up to 94% cheaper than alternatives.

  • Exa AI Search

    Exa Labs · Web Scraping

    Modern AI search engine with SERP API, website crawler, deep research API, and Websets for complex queries. Real-time web index updated every minute with SOC 2 Type II compliance.

  • OpenSandbox

    Alibaba · AI Infrastructure

    General-purpose sandbox platform for AI applications with multi-language SDKs, unified APIs, and Docker/Kubernetes runtimes. Built for coding agents, GUI agents, agent evaluation, and RL training.

  • WorldMonitor

    koala73 · AI Infrastructure

    Real-time global intelligence dashboard with AI-powered news aggregation, geopolitical monitoring, and infrastructure tracking in a unified situational awareness interface. 33K+ GitHub stars.

  • LLMfit

    AlexsJones · Inference Engine

    CLI tool to find which LLMs run on your hardware across hundreds of models and providers. One command to match model requirements to local GPU/CPU capabilities. 12K+ GitHub stars.

  • GitNexus

    abhigyanpatwari · AI Infrastructure

    Client-side knowledge graph engine that runs entirely in-browser. Drop in a GitHub repo or ZIP file to get an interactive knowledge graph with built-in Graph RAG agent for code exploration. 10K+ GitHub stars.

  • Auto-Round

    Intel · AI Infrastructure

    Accuracy-first quantization toolkit for LLMs supporting Weight-Only, MXFP4, NVFP4, GGUF, and adaptive schemes. Compatible with vLLM, SGLang, and Transformers for minimal quality degradation.

  • Paper to Notebook

    Paper2Notebook · Data Integration

    Converts research paper PDFs into executable PyTorch Jupyter notebooks using Gemini 2.0 Flash. Supports arXiv links and PDF uploads with trending paper discovery.

  • Alibaba Cloud Coding Plan

    Alibaba Cloud · Inference Engine

    Subscription API service for AI coding tools with flat-rate pricing. Compatible with Claude Code, OpenClaw, and Qwen Code. No surprise bills with predictable quota management.

  • Vibe2Real

    Vibe2Real · Evaluation & Benchmarks

    Debugging simulator with 15 real production failure scenarios training developers in incident response without AI assistance. 87% failure rate enforces genuine debugging competence.

  • e1s

    keidarcy · AI Infrastructure

    Terminal UI for managing AWS ECS resources — k9s-style interface for ECS services, tasks, and containers. Built in Go with ECS Exec support. 847 GitHub stars.

  • DreamServer

    Light Heart Labs · AI Infrastructure

    One-command local AI stack: LLM inference, chat UI, voice, agents, workflows, RAG, and image generation. Supports NVIDIA and AMD GPUs with Docker-based deployment. No cloud required.

  • SkillsMP

    SkillsMP · AI Infrastructure

    Agent Skills marketplace with 66K+ skills for Claude Code, Codex CLI, and ChatGPT. Community-curated SKILL.md files sourced from public GitHub repos with quality filtering, smart search, and category browsing.

  • memubot

    seigneurcui · AI Infrastructure

    Memory system for 24/7 proactive agents like OpenClaw (moltbot, clawdbot). Provides persistent conversation memory and context management for always-on AI agent deployments.

  • Daloopa Investing

    Daloopa · Data Integration

    Open-source investment research toolkit providing financial data extraction, analysis workflows, and AI-assisted due diligence for investors and analysts. 289 GitHub stars.

Research & Safety (4)

  • H-Neurons

    Tsinghua University · Interpretability

    Research identifying hallucination-associated neurons in LLMs. Less than 0.1% of total neurons reliably predict hallucination occurrences with strong cross-scenario generalization. Traces origins to pre-training phase.

  • Awesome AI Web Search

    felladrin · Survey & Overview

    Curated list of AI-powered web search tools covering search engines, metasearch, RAG implementations, and retrieval-augmented generation projects. 1.2K+ GitHub stars.

  • Agentic Reasoning Survey

    UIUC / Multiple Institutions · Survey & Overview

    Comprehensive survey on LLMs as autonomous agents capable of planning, acting, and learning. Organizes agentic reasoning into three layers: foundational capabilities, self-improvement via feedback, and multi-agent collaboration.

  • Open-Weight LLM Architecture Review

    Sebastian Raschka · Survey & Overview

    Deep-dive analysis of 10 open-weight LLM architectures from Jan–Feb 2026 including Trinity 400B, Kimi K2.5 1T, GLM-5 744B, Qwen3.5 397B, and Sarvam 105B. Covers MoE design, attention mechanisms, and efficiency trade-offs.

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