Quanta AI Labs

26 / Jun 22 – Jun 28, 2026 · 45 RELEASES

AI News Week 26, 2026

Custom Silicon + Agentic Trading. OpenAI previews GPT-5.6 Sol/Terra/Luna (Ultra subagent mode, 91.9% TerminalBench) and unveils Jalapeño — its first custom Broadcom-built inference ASIC (tape-out in ~9 months). IBM debuts the world's first sub-1nm (0.7nm) nanostack chip (~100B transistors). DeepReinforce ships Ornith 1.0 — self-scaffolding open coding models (9B→397B, 82.4 SWE-Bench Verified). ByteDance Seed 2.1 agent models + Seedance 2.x video. Alibaba Wan Streamer (25 FPS full-duplex audio-video world model), HappyHorse 1.0 (15B unified video+lip-synced audio, #1 Video Arena Elo), Qwen-AgentWorld (environment-modeling for agents). Krea 2 open weights (Raw + Turbo) + technical report. Google invests $75M in A24 for AI filmmaking tools. Sakana Fugu multi-agent orchestration API. Microsoft RD-Agent (automated R&D, leads MLE-bench). Huge agentic-trading wave: QuantDinger, OpenAlice, AgenticTrading (Open-Finance-Lab), LLM_trader, Claude-TradingView MCP. Research: NVIDIA ArtiFixer (3D artifact repair), 3D-DLP, Lift4D, DanceOPD flow distillation, UN-0 (oscillator-based image gen), PerceptionDLM (diffusion VLM), DomainShuttle, Meta RAM AutoData, BitRobot Humanoids-in-the-Wild 500h dataset, DifoTrain VLA. Frontier science: Aleph Neuro non-invasive ultrasound brain imaging, Midjourney Medical ultrasonic-CT scanner. Tools: vLLM Semantic Router, Claurst (Rust coding agent), Obsidian Copilot, CasaOS, AI Observer, Facebook Ads Library MCP, Loop Engineering, mattpocock/skills, Mosaic video AI, OpenArt Director, Leaf Node agriculture hardware.

Language Models (5)

  • GLM-5.2 Local Guide

    Linas (Substack) · Text Generation

    Guide to running Zhipu AI's open-weight GLM-5.2 (MIT) locally — 1M-token context, runs on a single 256GB Mac via dynamic quantization, ranks #1 for frontend coding on Arena AI, rivaling Claude Opus and GPT-5.5.

  • Ornith 1.0 (HF Collection)

    DeepReinforce · Code Generation

    Open-source (MIT) agentic coding model family spanning 9B Dense to 397B MoE that uses RL to learn not just solution rollouts but the scaffolds driving them; the 397B scores 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1, the 9B 69.4 SWE-Bench Verified.

  • Ornith 1.0

    DeepReinforce · Code Generation

    Self-scaffolding open-source coding model family (9B/31B Dense, 35B/397B MoE, MIT license) where RL jointly learns solution rollouts and the scaffolds that drive them; the 397B reaches 82.4 SWE-Bench Verified (second only to Claude Opus 4.8's 87.6) and 77.5 Terminal-Bench 2.1.

  • ByteDance Seed 2.1

    ByteDance · Code Generation

    Multimodal AI agent model family (Pro and Turbo) for real-world productivity tasks including code engineering, office work and complex reasoning; scores 87.0 on BeyondAIMER, 92.6 on MathVision (with tools) and 89.2 on VideoMME.

  • GPT-5.6 Sol

    OpenAI · Reasoning

    Next-gen model family previewed June 26 in three variants — Sol (flagship), Terra (balanced), Luna (affordable) — with a new max reasoning effort and a subagent-driven Ultra mode; Sol scores 88.8% (Ultra 91.9%) on TerminalBench 2.1 and ~9 points above GPT-5.5 on biology evals, priced $5/$30 per 1M tokens, initially limited to ~20 orgs.

Vision & Image (4)

  • Krea 2 Open Weights

    Krea AI · Text-to-Image

    Krea announced open weights for Krea 2 in two variants — Krea 2 Raw (undistilled mid-training model for fine-tuning) and Krea 2 Turbo (distilled, ~2-second generation) — released on Hugging Face under a custom license with enterprise terms for 50+ seats and mandatory safety safeguards.

  • DanceOPD

    DanceOPD · Text-to-Image

    Flow-matching distillation framework that merges multiple generative capabilities (text-to-image, local and global editing) into one model by treating each source capability as a velocity field, hard-routing one frozen field per sample, and training the student via velocity MSE on student-visited states.

  • UN-0

    unconv.ai · Text-to-Image

    Image generator built on a simulated system of coupled Kuramoto oscillators instead of neural network layers, letting oscillators evolve freely from random phases and decoding only the final snapshot at time T; achieves FID 6.74 on ImageNet 64x64.

  • PerceptionDLM

    MSA Lab, Peking University · Visual Understanding

    Multimodal diffusion language model (LLaDA-8B-Instruct backbone + vision encoder) that generates region descriptions in parallel within a single denoising pass; outperforms LLaDA-V on 15/16 benchmarks and hits 62.4% on ParaDLC-Bench in 276s vs 479s for autoregressive baselines.

Video & Animation (8)

  • Mosaic

    Mosaic · Video Editing

    Agentic AI platform for video editing that uses AI agents to automate and assist editing tasks. Deeper architecture details are not disclosed on the landing page.

  • Happy Oyster

    Happy Oyster · Text-to-Video

    Real-time interactive world model that generates coherent audio and video simultaneously with physics consistency, letting users direct scenes via text, voice or images or explore procedurally generated first-person worlds.

  • Seedance 2.x

    ByteDance · Text-to-Video

    Video generation update (per xiaohu) producing 30-second clips in a single generation with native 4K resolution and up to 50 full-mode reference inputs plus 3D model support, launched alongside an AI copyright commercialization platform for licensed movie IP with revenue sharing.

  • Google x A24 AI Filmmaking

    Google DeepMind / A24 · Text-to-Video

    Google is investing $75M in studio A24 via a DeepMind partnership to build AI filmmaking tools, focused on generating storyboards and new creative workflows that preserve filmmaker control rather than cutting costs.

  • OpenArt Director

    OpenArt · Text-to-Video

    Conversational AI video creation tool that generates videos from chat prompts, integrating image generation, video editing, lip-sync, motion sync and voice-over audio.

  • Wan Streamer v0.1

    Alibaba Wan Team · Real-time Video

    Native-streaming end-to-end interactive foundation model that unifies language, audio and video as interleaved input/output tokens in a single Transformer with block-causal attention, running at 25 FPS with ~200 ms model-side latency and sub-second full-duplex audio-visual interaction.

  • DomainShuttle

    cn-makers · Text-to-Video

    Subject-driven text-to-video system for cross-domain personalization that maps real-world subjects across diverse visual domains (fantasy, illustration styles) while preserving identity consistency, supporting in-domain and open-domain creative generation with text control.

  • maxescu — Shared Post

    Alex Patrascu (@maxescu) · Text-to-Video

    Shared AI product/demo post from generative-AI curator @maxescu — see the linked X post for the prompt and full details.

3D & Spatial (6)

  • ArtiFixer

    NVIDIA Research · 3D Generation

    Two-stage generative pipeline that fixes 3D reconstruction artifacts in under-observed regions by finetuning a bidirectional video diffusion model with an opacity-mixing strategy, then distilling it into a causal autoregressive model generating hundreds of frames per pass; 1-3 dB PSNR gains over SOTA on MipNeRF 360 and DL3DV-10K.

  • 3D-DLP

    eubooks3003 · 3D Reconstruction

    Self-supervised 3D object-centric scene representation that decomposes scenes into interpretable latent particles (position, bbox, appearance) from RGB-D/voxel input via a K-means prior and 3D volumetric decoder; 74.5% on RLBench (+7.3 vs 2D) and 48.1% on MimicGen (+14.0).

  • Midjourney Medical (Blog)

    Midjourney · 3D Reconstruction

    Midjourney's first hardware product: a full-body Ultrasonic CT scanner that lowers a person into a shallow pool ringed with ~500,000 ultrasonic sensors, firing soundwaves from every angle and reconstructing them into 3D cross-sections of muscle, fat, bone and organs in a 60-second scan.

  • Arbor (3D Control)

    jdihlmann · 3D Generation

    Controlled 3D asset generation method that fuses typed constraint meshes (hull, avoidance, touch regions) into one surface signal, encoding them via frozen encoders into compact tokens routed through a diffusion model's geometry branch for precise spatial control without retraining the base 3D generator.

  • Lift4D

    Lift4D · 3D Reconstruction

    Test-time optimization framework that reconstructs dynamic 3D objects from a single monocular video by adapting an existing single-view 3D reconstruction model for temporal consistency with deformable 3D Gaussian Splatting, refined via occlusion-aware optimization and diffusion-based inpainting for unobserved regions.

  • Aleph Ultrasound Brain

    Aleph Neuro · 3D Reconstruction

    Non-invasive brain imaging technology that sends ultrasound waves through the intact skull and reads scattering off red blood cells to map neural activity via blood flow, reportedly achieving MRI-comparable resolution without surgical implants or bulky equipment.

Agents & Automation (12)

  • Claude TradingView MCP Trading

    jackson-video-resources · Agent Orchestration

    Node.js bot that connects Claude to TradingView charts and crypto exchanges (BitGet, Binance) via Model Context Protocol to auto-execute trades from JSON-defined strategy rules, with paper-trading mode, safety checks and CSV tax logging.

  • AgenticTrading

    Open-Finance-Lab · Agent Framework

    Open-source platform for building and testing LLM-powered trading agents via backtesting and live paper trading, combining a FastAPI + SQLite backend with the FinAgent orchestration framework and a dashboard for inspecting agent reasoning logs and performance.

  • LLM Trader

    qrak · Agent Framework

    Autonomous crypto trading agent pairing vision-based chart analysis with memory-augmented reasoning, using Gemini (with OpenRouter/LM Studio fallbacks) and a vector DB of past trades; runs paper-trading only with no live exchange integration.

  • DifoTrain

    AshrafGalibShaik · Robotics

    Embodiment-agnostic framework for training Vision-Language-Action (VLA) models from real-world demonstrations, supporting teleoperation/human-video data collection, behavior-cloning training on standardized datasets, and pluggable robot, data and policy interfaces for sim or real deployment.

  • OpenAlice

    TraderAlice · Agent Framework

    AI trading agent in a TypeScript/Python monorepo with a two-process design: an Alice runtime holding agent logic/research tools and a separate UTA service for broker connections and execution; runs natively in Claude Code or other CLI tools via MCP, with all trades requiring user approval.

  • Obsidian Copilot

    Logan Yang · Agent Framework

    Obsidian plugin implementing a RAG system over the user's vault with multi-provider LLM support and an agentic layer for autonomous tool calling across notes, web search and multimedia processing.

  • Claurst

    Kuberwastaken · Coding Agents

    Open-source, multi-provider terminal coding agent written in Rust — a TUI pair programmer implementing the Agent Client Protocol (ACP) for editor integration via a clean-room reimplementation of Claude Code behavior, running fully locally with no telemetry.

  • Qwen-AgentWorld

    Alibaba · Agent Framework

    Two models (released ~June 24, 2026) trained to predict what agent environments return across seven domains — MCP, Search, Terminal, Software Engineering, Android, Web and OS — enabling environment-modeling for agent training.

  • Loop Engineering

    Cobus Greyling · Agent Framework

    Reference implementation and pattern library for recursive automation systems that orchestrate AI coding agents over time, with CLI tools (loop-audit, loop-init, loop-cost) and starters for Grok, Claude Code and Codex covering scheduling, state management, sub-agents and verification loops.

  • Humanoids in the Wild (500h)

    BitRobot Foundation · Robotics

    Robot learning dataset with 500+ hours of human-teleoperated Unitree G1 humanoid data performing household tasks in real Southeast Asian homes, with multimodal streams (RGB/IR stereo, joint states, IMU) across 11 task categories and fine-grained subtask annotations for mobile manipulation.

  • Meta RAM AutoData

    Meta (FAIR) · Agent Framework

    Framework enabling AI agents to act as data scientists that iteratively build high-quality training and evaluation data, generating, analyzing and refining synthetic datasets through feedback loops between weak and strong model solvers.

  • Sakana Fugu

    Sakana AI · Agent Orchestration

    Multi-agent orchestration system delivered as a single API that dynamically assembles, routes and coordinates expert agents per task (based on the TRINITY and Conductor ICLR 2026 papers), charging a single blended rate; Fugu and Fugu Ultra match or exceed GPT-5.5, Gemini 3.1 Pro and Opus 4.8 across coding, reasoning and science benchmarks.

Developer Tools (9)

  • Pocock Skills

    Matt Pocock · Prompt Engineering

    Collection of reusable Claude Code skills (structured prompts/workflows) targeting common agent failure modes like verbosity, developer-agent misalignment and architectural degradation, covering TDD, domain modeling and codebase architecture.

  • vLLM Semantic Router

    vLLM Project · Inference Engine

    Signal-driven intelligent router that dynamically selects among multiple LLMs based on workload characteristics, optimizing for token efficiency, safety detection and cost across cloud, datacenter and edge deployments.

  • AI Observer

    tobilg · AI Infrastructure

    Self-hosted observability backend (Go + React, DuckDB storage) that aggregates telemetry from multiple AI coding assistants into one dashboard, ingesting via OTLP on port 4318 and serving a REST API and WebSocket hub on port 8080.

  • LLM-From-Scratch

    Thoyzvamsi · Inference Engine

    Educational repo implementing a character-level GPT-style transformer (6 transformer blocks with multi-head self-attention and feedforward layers) trained on Wikipedia text, with a full data pipeline for scraping and tokenization.

  • CasaOS

    IceWhale Tech · AI Infrastructure

    Open-source personal cloud OS (Go + Vue.js) that turns consumer hardware into a self-hosted Docker-based app and file server, supporting x86-64, ARM64 and ARMv7 on Linux.

  • Facebook Ads Library MCP

    Ramses Aguirre · AI Infrastructure

    FastMCP-based Model Context Protocol server with 15+ tools wrapping the Facebook Graph API for competitive ad intelligence — analyzing competitor ad strategies, tracking campaign changes and predicting ad performance — exposed directly to Claude.

  • Leaf Node v3.x

    LeafAI · Edge Deployment

    Open-source ESP32-S3-based environmental monitoring/control device for precision agriculture that reads sensors over I2C, RS485, SDI-12 and OneWire, publishes readings via MQTT and triggers actuators like lights, fans and pumps.

  • Jalapeño Inference Chip

    OpenAI / Broadcom · Inference Engine

    OpenAI's first custom chip — a reticle-sized ASIC built from scratch for LLM inference (not a repurposed accelerator) that delivers substantially better performance-per-watt by minimizing data movement; co-developed to tape-out in ~9 months with help from OpenAI's own models, deployment starting late 2026.

  • IBM Sub-1nm Chip

    IBM · AI Infrastructure

    World's first sub-1nm (0.7nm node) chip technology using a 'nanostack' architecture that vertically stacks and staggers transistors in 3D, packing nearly 100 billion transistors on a fingernail-sized chip with 40% SRAM scaling and up to 50% more performance or 70% better energy efficiency vs IBM's 2nm chips.

Research & Safety (1)

  • Krea 2 Technical Report

    Krea AI · Research Paper

    Technical report for Krea 2, an open-weights text-to-image diffusion transformer using GQA with gated sigmoid attention and Qwen 3 VL text encoding, with a multi-stage pretraining-through-RL pipeline on PyTorch and data curation that prioritizes stylistic diversity over conventional quality filtering.

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