Quanta AI Labs

19 / May 4 – May 10, 2026 · 73 RELEASES

AI News Week 19, 2026

Frontier and creative push: Google ships Gemma 4 31B IT Assistant + Multi-Token Prediction drafters (3x faster Gemma 4) + Gemini file generation + AlphaEvolve (Gemini-powered algorithm discovery) + Vision Banana. Zyphra releases ZAYA1 8B. Baidu launches ERNIE 5.1 (2.4T native full-modality powering Wenxin Yiyan). antirez builds a custom Metal/CUDA DS4 inference engine for DeepSeek V4 Flash on 96–128GB Macs. ElevenLabs ships ElevenMusic. Kling Video 3.0 adds 4K mode. Magnific One relaunches (Freepik rebrand). Topaz Labs Next-Gen update. Bach-1.0 Preview debuts at #6 on T2V leaderboard. Luma updates Dream Machine. UniVidX, SwiftI2V, Quant VideoGen, and CDM advance video research. NVIDIA ships D-Rex relightable digital humans, MotionBricks for humanoid control on G1, and open-sources Audio2Face-3D. Meta open-sources MapAnything (universal metric 3D). Ai2 releases MolmoAct 2 (37x faster than v1). Anthropic ships Claude for Financial Services. OpenAI launches GPT-Realtime-2 / Translate / Whisper voice models. Databricks Genie. Cursor publishes their agent-harness optimization playbook. Capy and Terminal-Bench redefine coding-agent infra. Tooling explosion: ForgeCode, Free Claude Code, Octogent, Agent Lattice (lat.md), mcpc, liter-llm, Obsidian Skills, Postgres LLM, Scrapling, Three Man Team, OpenMontage, Tolaria, Build-Your-Own-OpenClaw, and Karpathy-inspired Context Engineering handbook. ProgramBench (Meta FAIR) measures whether agents can rebuild programs from scratch.

Language Models (6)

  • Gemma 4 31B IT Assistant

    Google · Text Generation

    Google's instruction-tuned Gemma 4 31B model fine-tuned as a general-purpose assistant variant, continuing the Gemma open-weights line with stronger conversational and reasoning behavior.

  • ZAYA1 8B

    Zyphra · Small Language Models

    Zyphra releases ZAYA1 8B, an 8B open-weights model built on their hybrid SSM-Transformer architecture, targeting efficient long-context inference at small scale.

  • Gemini File Generation

    Google · Text Generation

    Gemini can now natively generate downloadable files (DOCX, PDF, spreadsheets) directly from chat, turning the assistant into an end-to-end document authoring tool.

  • ERNIE 5.1 (Wenxin Yiyan)

    Baidu · Text Generation

    Baidu's flagship 2.4T-parameter native full-modality model powering Wenxin Yiyan, with unified text/image/audio/video generation and agentic capabilities surpassing DeepSeek-V4-Pro.

  • Sparser Faster Lighter Transformer LMs

    Sakana AI & NVIDIA · Text Generation

    ICML 2026 paper using L1 regularization to induce 99%+ feedforward sparsity, with new CUDA kernels delivering up to 24% training speedups and 24%+ peak GPU memory reduction.

  • Multi-Token Prediction in Gemma 4

    Google / Hugging Face · Text Generation

    MTP drafters deliver up to 3x inference speedup on Gemma 4 via speculative decoding that shares KV cache with the target model, with drafters available for E2B, E4B, 26B A4B, and 31B.

Vision & Image (5)

  • HiDream O1 Image

    HiDream AI · Text-to-Image

    HiDream releases its O1 Image model on Hugging Face, an open-weights text-to-image generator continuing HiDream's high-fidelity diffusion lineup.

  • Vision Banana

    Google DeepMind · Image Editing

    Vision Banana extends the Nano Banana family into a vision-focused image generation and editing model with stronger instruction following and visual reasoning.

  • Magnific One

    Magnific / Freepik · Image Editing

    Freepik's Magnific relaunches as 'Magnific One,' consolidating its image upscaling, relighting, and generative editing tools into a single creative platform for pros.

  • Topaz Labs Next-Gen Update

    Topaz Labs · Image Restoration

    Next-Gen update with improved AI models for photo and video enhancement — sharpening, denoising, and upscaling at professional grade.

  • WenWare

    WenWare · Text-to-Image

    AI-powered GeoGuessr clone where players guess both the location and the year of AI-generated 360 panoramas, built solo with GPT-Image-2 and Codex.

Video & Animation (7)

  • Quant VideoGen

    SVG Project · Text-to-Video

    Training-free 2-bit KV-cache quantization framework for autoregressive long video generation, cutting KV memory up to 7x with <4% latency overhead while improving quality.

  • MotionBricks

    NVIDIA · Real-time Video

    Modular latent generative model with smart primitives for scalable real-time motion synthesis, used for interactive animation and humanoid robot control on Unitree G1.

  • Bach-1.0 Preview

    Video Rebirth · Text-to-Video

    Bach-1.0 Preview is a new text-to-video model that debuted at #6 on the Artificial Analysis T2V leaderboard, signaling a competitive new entrant in video generation.

  • Luma Dream Machine Update

    Luma Labs · Text-to-Video

    Luma repositions as 'creative agents that make you prolific,' rolling out a Dream Machine update emphasizing directable intelligence and ship-ready aesthetic for video and image gen.

  • UniVidX

    SIGGRAPH 2026 · Text-to-Video

    Omni-directional video generation model that synthesizes 360° panoramic video, enabling fully immersive VR-ready content from prompts.

  • SwiftI2V

    SwiftI2V · Text-to-Video

    Efficient high-resolution image-to-video generation that splits output into conditional segments and generates them progressively, slashing compute vs. one-shot approaches.

  • Grainrad

    Grainrad · Video Editing

    Free WebGPU-powered browser playground that turns images and videos into live 60fps effects like ASCII art, halftone, pixel sorting, VHS, and matrix rain.

Audio & Speech (3)

  • ElevenMusic

    ElevenLabs · Music Generation

    ElevenLabs launches ElevenMusic, an AI music generator producing full songs with vocals, instrumentals, lyrics, and rap beats from text prompts.

  • Overwhisper

    OverseedAI · Speech-to-Text

    Native macOS menu bar app for voice transcription using WhisperKit on-device with optional OpenAI Whisper cloud fallback, global hotkeys, and direct cursor text insertion.

  • GPT-Realtime-2, Translate & Whisper

    OpenAI · Voice Agents

    Three new realtime voice models in the API: GPT-Realtime-2 with GPT-5-class reasoning, GPT-Realtime-Translate for 70+ language live translation, and GPT-Realtime-Whisper for streaming speech-to-text.

3D & Spatial (7)

  • D-Rex

    NVIDIA Research · 3D Generation

    D-Rex generates photorealistic, relightable digital humans by combining the EVA avatar model with Cosmos DiffusionRenderer for production-grade synthetic performers.

  • Stepper

    Stepper Scene · 3D Generation

    Stepwise immersive scene generation system that builds explorable 3D environments from multiview panoramas, growing scenes incrementally rather than in one shot.

  • RecGen

    RecGen Project · 3D Reconstruction

    Reconstructive generation from RGB-D input, fusing depth-aware reconstruction with generative completion to produce coherent 3D scenes from partial observations.

  • PhysForge

    PhysForge · 3D Generation

    Generates physics-grounded 3D assets ready for interactive virtual worlds — objects with consistent mass, collision, and material behavior baked into the generation process.

  • Map2World

    Map2World · 3D Generation

    Turns a 2D segmentation map plus text prompt into a complete 3D world, conditioning scene generation on layout maps for controllable environment design.

  • MapAnything

    Meta · 3D Reconstruction

    Meta and CMU's universal feed-forward model for metric 3D reconstruction from any image input, producing scale-accurate geometry in a single forward pass.

  • Audio2Face-3D

    NVIDIA · 3D Generation

    NVIDIA open-sources its Audio2Face-3D collection — pretrained models, MIT-licensed C++/Python SDK, Maya and Unreal 5 plugins, and Apache-licensed training framework for audio-driven facial animation.

Agents & Automation (24)

  • AlphaEvolve

    Google DeepMind · Coding Agents

    Gemini-powered coding agent that uses evolutionary search to discover advanced algorithms, already producing improvements across math, chip design, and data-center scheduling.

  • Claude Code Harness

    Chachamaru127 · Coding Agents

    Development harness that turns Claude Code into a disciplined Plan-Work-Review autonomous cycle with a Go-native core and a 5-verb skill system.

  • OpenDirectory

    Varnan-Tech · Agent Framework

    54 pre-built AI agent skills for GTM, technical marketing, and growth automation that work with Claude, Codex, and Gemini CLI in seconds with no setup.

  • Noustiny

    UfukNode · Coding Agents

    Agent-native video creation pipeline on Hermes Agent: drop a story seed, walk branching what-ifs, then auto-render visuals, cloned narration, and a shareable video.

  • AutoHedge

    The Swarm Corporation · Agent Framework

    Autonomous agent hedge fund using swarm intelligence with Director, Quant, Risk, and Execution agents for end-to-end market analysis and trade execution on Solana.

  • Three Man Team

    russelleNVy · Agent Orchestration

    Structured 3-agent dev process (Architect, Builder, Reviewer) with token-optimized handoffs that prevents AI coding drift; works with Claude Code, Cursor, and VS Code.

  • Octogent

    hesamsheikh · Agent Orchestration

    Thin orchestration dashboard over Claude Code that gives each parallel agent its own scoped context, notes, and task list, and lets Claude spawn sub-agents.

  • Agent Lattice (lat.md)

    Yury Selivanov · Agent Framework

    A markdown knowledge graph for your codebase that replaces flat AGENTS.md files with wiki-linked sections, code backlinks, and a drift checker.

  • Obsidian Skills

    kepano · Agent Framework

    Agent skills for Obsidian following the Agent Skills spec that teach Claude Code, Codex, and others to use Markdown, Bases, JSON Canvas, and the Obsidian CLI.

  • Higgsfield Skills

    Higgsfield AI · Agent Framework

    Official Higgsfield agent skills for image and video generation including Marketing Studio and Virality Predictor scoring; works with Claude Code, Cursor, and Codex.

  • ATLAS

    General Intelligence Capital · Agent Framework

    Self-improving AI trading agents built with Karpathy-style autoresearch; ships live signals, plain-English strategy builder, and a copy-trading marketplace.

  • Claude for Financial Services

    Anthropic · Agent Framework

    Anthropic's reference agents, skills, and connectors for investment banking, equity research, PE, and wealth management, deployable as Cowork or Managed Agents.

  • TinyFish Web Agent Integrations

    TinyFish · Web Agents

    Integrations connecting TinyFish's natural-language web automation agent to workflow tools and AI agent frameworks, starting with a Dify plugin.

  • OpenMontage

    calesthio · Agent Framework

    Open-source agentic video production system with 12 pipelines, 52 tools, and 500+ agent skills that turns any AI coding assistant into a video production studio.

  • MolmoAct 2

    Allen Institute for AI (Ai2) · Robotics

    Open robotics foundation model with faster 3D action reasoning, running 37x faster than its predecessor and ranking first on 7 of 8 third-party manipulation benchmarks. Includes 720-hour bimanual dataset.

  • Databricks Genie

    Databricks · Agent Framework

    State-of-the-art data agent using specialized knowledge search, parallel thinking, and multi-LLM design to answer complex enterprise data questions, lifting internal benchmark accuracy from 32% to over 90%.

  • ForgeCode

    Tailcall · Coding Agents

    AI-enabled terminal pair programmer with three specialized agents (forge, sage, muse), semantic codebase search, MCP and skills support, and access to 300+ models across Claude, GPT, Gemini, Grok, and DeepSeek.

  • Capy Open Source Program

    Capy · Coding Agents

    Free access to Capy's AI-native cloud IDE for open-source projects, running up to 25 concurrent coding agents in sandboxed VMs with Claude, GPT-5, Gemini, Grok, Kimi K2, and Qwen.

  • Cursor: Continually Improving Our Agent Harness

    Cursor · Agent Framework

    Cursor details how it tunes its agent harness per model, measures via CursorBench and Keep Rate, and stacks small optimizations to make agents noticeably faster and smarter inside their environment.

  • GENE-26.5

    Genesis AI · Robotics

    First AI brain for human-level robotic manipulation, paired with proprietary tactile gloves 100x cheaper than alternatives. Demos include 20-step cooking, smoothies, Rubik's cube, and piano.

  • LabOS

    Zaixi Zhang et al. · Agent Framework

    AI-XR co-scientist combining multi-agent reasoning, a self-evolving 'Tool Ocean', and AR smart-glass guidance for lab work, with LabSuperVision VLM benchmark for biomedical and materials labs.

  • The Solo Content Studio

    The Solo Content Studio · Agent Framework

    Platform that gives solo creators ten trained AI specialists in their own voice, running on a schedule so one person can ship content like a ten-person studio.

  • Agent Exchange (AEX)

    open-experiments · Agent Framework

    Programmatic marketplace ('NASDAQ for AI Agents') brokering discovery, bidding, contract award, and settlement between consumer and provider agents via the AWE protocol with A2A direct comms.

  • Forge NPS v01

    0xzgbot · Agent Framework

    Cinematic content production orchestrator led by the Hermes agent with Kimi Moonshot planner, a 127-skill library, persistent memory, and multi-model rendering across Flux2, ComfyUI, and Spark.

Developer Tools (15)

  • DS4 (DeepSeek 4 Flash Inference)

    Salvatore Sanfilippo (antirez) · Inference Engine

    antirez's DS4 is a dedicated Metal/CUDA inference engine for DeepSeek V4 Flash (284B params, 1M context), exploiting compressed KV cache + on-disk persistence to run on 96–128GB Macs.

  • Hunyuan3D 2.0/2.1 Windows Portable

    YanWenKun · Edge Deployment

    Portable Windows pack for running Tencent's Hunyuan3D 2.0/2.1 Gradio demo on consumer GPUs via mmgp offloading — needs only 3GB VRAM for geometry, 6GB for textures.

  • GPT Image 2 Skill

    wuyoscar · Prompt Engineering

    Prompt gallery, agentic skill, and CLI for OpenAI GPT Image 2 with curated copy-paste prompts and runnable examples for skill-capable agents.

  • Build Your Own OpenClaw

    czl9707 · AI Infrastructure

    18-step progressive tutorial to build a minimal OpenClaw-style AI agent from a chat loop to a full agent with tools, skills, memory, and web access.

  • TALOS-V2

    Luthiraa · Inference Engine

    SystemVerilog FPGA implementation of Karpathy's microGPT on DE1-SoC/Cyclone V running transformer inference at 50k+ tokens per second over JTAG.

  • Tolaria

    refactoringhq · AI Infrastructure

    Files-first, git-first, offline desktop app for managing markdown knowledge bases as second brains, company docs, or OpenClaw/agent memory and procedures.

  • dLLM

    ZHZisZZ · Evaluation & Benchmarks

    Library that unifies training and evaluation of diffusion language models for transparent, reproducible end-to-end development of dLLMs.

  • Upwork MCP

    vanooo · AI Infrastructure

    MCP server using Chrome DevTools Protocol to let Claude Code search Upwork jobs and manage proposals, messages, and contracts without tripping bot detection.

  • liter-llm

    kreuzberg-dev · Inference Engine

    Universal LLM API client with a Rust core powering 142+ providers across 11 native language bindings including Python, Node, Java, Go, C#, PHP, and WASM.

  • mcpc

    Apify · AI Infrastructure

    Universal MCP CLI client with persistent sessions, stdio/HTTP, OAuth 2.1, async tasks, JSON code-mode output, AI sandboxing proxy, and x402 agentic payments.

  • Context Engineering

    davidkimai · Prompt Engineering

    Karpathy-inspired first-principles handbook moving beyond prompt engineering to the wider discipline of context design, orchestration, and optimization.

  • Postgres LLM

    JigsawStack · AI Infrastructure

    Dynamic Postgres trigger function that runs async LLM requests directly inside the database for translation, classification, summarization, sentiment, image analysis, and web search.

  • Claude Code Guide

    Cranot · Prompt Engineering

    Complete Claude Code CLI reference auto-updated every 2 days from official docs, GitHub releases, and Anthropic changelogs; optimized for both humans and agents.

  • PDFCraft

    PDFCraft · Edge Deployment

    Free MIT-licensed PDF toolkit with 90+ tools that runs entirely in-browser via WebAssembly, pdf-lib, pdf.js, and tesseract.js, with a drag-and-drop workflow editor.

  • agent-skills-eval

    darkrishabh · Evaluation & Benchmarks

    Test runner that empirically validates agentskills.io-style skills by running prompts with and without the skill loaded and judge-grading both outputs to measure real improvement.

Research & Safety (6)

  • Zyphra

    Zyphra · Survey & Overview

    Open-frontier AI lab building hybrid state-space + transformer foundation models with the tagline 'The future of intelligence is open.'

  • CDM (Continuous-time Distribution Matching)

    CDM Diffusion · Research Paper

    Few-step diffusion distillation method that compresses diffusion samplers into 1–4 step generators while preserving full quality.

  • ProgramBench

    Meta FAIR · Benchmark

    Benchmark of 200 tasks where SWE-agents must rebuild full programs (FFmpeg, SQLite, PHP) from only an executable and documentation, evaluated via agent-driven behavioral fuzzing.

  • Terminal-Bench

    Stanford & Laude Institute · Benchmark

    Benchmark suite for AI agents in terminal environments, spanning software engineering, security, system administration, and data science tasks across versions 1.0-3.0.

  • Startups.RIP

    Startups.RIP · Dataset

    Database of 1,700+ defunct or acquired YC startups from 2005 onward, with retros and rebuild playbooks framing dead companies as lessons rather than failures.

  • Evaluating AI Agent Skills

    Langfuse · Benchmark

    Methodology for evaluating Claude-style agent skills like prompts: store user prompts as a Langfuse dataset, run agents, trace behavior, and LLM-as-judge score the output.

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