Quanta AI Labs

24 / Jun 8 – Jun 14, 2026 · 48 RELEASES

AI News Week 24, 2026

Frontier Flagships + Google I/O 2026. Anthropic ships Claude Fable 5 (new flagship — SOTA coding/finance/vision, million-token reasoning) + Mythos 5 (safeguard-relaxed variant for vetted cyber researchers), $10/$50 per M tokens. Google I/O lands Gemini Omni (any-input generative, video first), Gemini 3.5 Flash (GA via Antigravity), Gemini Live 3.5 Translate (continuous speech-to-speech, 70+ languages, 2000+ pairs), and DiffusionGemma (26B MoE diffusion text gen, 1000+ tok/s on H100). Z.ai GLM-5.2 (744B MoE, 1M context, dual thinking-effort). MiniMax M3 (428B/23B-active native multimodal, 1M context) + open MiniMax Sparse Attention kernels. Nex AGI Nex-N2 (397B MoE agentic reasoning, Apache 2.0). Moonshot Kimi Code (K2.7). Creative wave: Luma Ray3.2 (16 keyframes, HDR 1080p, Modify Video V2), Runway Gen-4.5 (tops Video Arena), ElevenLabs Music v2 (genre transitions, licensed), ByteDance Bernini-R video renderer, Z.ai SCAIL-2 character animation, Tripo Project Eden persistent world model. TTS: RedNote dots.tts (continuous AR, 24 langs), Miso-TTS (110ms latency voice agents). Massive 3D/world-model research: NVIDIA Locate Anything (parallel box decoding), reshoot-anything, WorldString neural twins, OSCAR robot world model, StreamForce force-driven video, World Tracing image-to-3D, VideoMDM, SurFlo, MoVerse, Princeton i1 open T2I, AnchorWorld, MeshFlow, MilliVid long-video, Flex4DHuman. Benchmarks: Agents' Last Exam (55 sub-industries), Arbor research agent (hypothesis-tree). Tools: ComfyUI comfy-cli, CLI Printing Press, OpenMed clinical NLP, PriceGhost, Shotblock 3D storyboards, Get It study app, train-llm-from-scratch. Industry: Microsoft AI Playground (MAI models), YouTube prominent AI labels, Prometheus AI studio. Plus shared community workflows from fofr, and creator demos.

Language Models (6)

  • Claude Fable 5 & Mythos 5

    Anthropic · Text Generation

    Anthropic's new flagship: Fable 5 is the general-purpose release (SOTA coding, finance, vision, long-context reasoning over millions of tokens); Mythos 5 is the same model with safeguards relaxed for vetted cybersecurity researchers. $10/M input, $50/M output, with new cyber + bio/chem classifiers and distillation protection that falls back to Opus 4.8 when triggered.

  • Gemini Omni & Gemini 3.5 Flash

    Google · Text Generation

    Google I/O 2026 model announcement: Gemini Omni is an any-input generative model starting with video, emphasizing world understanding, multimodality and in-chat editing; Gemini 3.5 Flash is a frontier intelligence-plus-action model, GA via Google Antigravity, the Gemini API, AI Studio and Android Studio.

  • DiffusionGemma

    Google · Text Generation

    Experimental Apache-2.0 diffusion-based text-generation model (26B MoE, 3.8B active, ~18GB VRAM quantized) generating 256-token blocks in parallel via iterative refinement with bidirectional attention — 1000+ tok/s on H100, ~700 on RTX 5090 (up to 4x faster than autoregressive). Built on Gemma 4 + Gemini Diffusion research; excels at speed-critical infilling/inline edits.

  • MiniMax M3

    MiniMax AI · Text Generation

    Native-multimodal MoE LLM (~428B total / ~23B active) with a 1M-token context, trained mixed-modality (text/image/video) from step one and using MiniMax Sparse Attention for 9x prefill / 15x decode speedups over M2 at 1M context (per-token compute cut to ~1/20). Thinking and non-thinking modes, frontier long-horizon agentic + coding performance.

  • Nex-N2

    Nex AGI · Text Generation

    Open-source (Apache 2.0) agentic "thinking" model family that adaptively decides when and how deeply to reason, unifying coding, search and tool use into one closed loop. Nex-N2-Pro is a 397B MoE (17B active) post-trained on Qwen3.5-397B-A17B; a Mini variant is 35B MoE (3B active), with free launch-window API access on OpenRouter and SiliconFlow.

  • GLM-5.2

    Z.ai (Zhipu) · Text Generation

    Z.ai's June 13 flagship agentic-coding model: same 744B-param MoE architecture as GLM-5 with a usable 1M-token context (glm-5.2[1m]) and up to 131K output tokens, adding a dual thinking-effort system (High and Max). Launched across all GLM Coding Plan tiers, with standalone API, chatbot and MIT open weights to follow.

Vision & Image (4)

  • Locate Anything (NVIDIA Research)

    NVIDIA · Visual Understanding

    Unified vision-language grounding/detection model (Moon-ViT encoder + Qwen2.5 decoder) using Parallel Box Decoding to predict full bounding boxes in one step instead of token-by-token. Trained on 138M queries / 785M boxes — 12.7 boxes/sec (>10x faster than Qwen3-VL), +3.8% mean F1 on LVIS, SOTA on ScreenSpot-Pro (60.3) and M6Doc (70.1).

  • Locate Anything — Live Demo

    NVIDIA · Visual Understanding

    Interactive Hugging Face Space for NVIDIA's Locate Anything — try the parallel-box-decoding grounding model in-browser on your own images for detection, GUI/document localization and dense grounding.

  • fofr — Community Workflow

    fofr (@fofrAI) · Text-to-Image

    Shared workflow/demo from fofr (Replicate creator known for Flux fine-tunes and ComfyUI pipelines). See the linked X post for the full prompt and outputs.

  • i1

    Princeton (Z-Lab) · Text-to-Image

    Fully open 3B-param text-to-image diffusion model with transparent design, data and code: dual-stream MMDiT with long skip connections, a single T5Gemma-2B text encoder + 2-block adapter, trained progressively at 256/512/1024 over 700K+ TPU v6e hours on 12 public datasets with Qwen3-VL synthetic captions. Beats the best prior fully-open model by 29.5 absolute points avg across five benchmarks.

Video & Animation (10)

  • SCAIL-2

    Z.ai (zai-org) · Video Editing

    14B end-to-end character-animation model built on Wan 2.1 that transfers motion from driving videos to reference characters with no intermediate pose/skeleton, using a Unified Motion Transfer Interface with two mask channels, mode-specific RoPE and Bias-Aware DPO. Trained on synthetic MotionPair-60K; supports single/multi-character animation, character replacement, zero-shot animal driving at 512p/704p.

  • Bernini-R

    ByteDance · Text-to-Video

    DiT-based renderer for ByteDance's Bernini video framework (Latent Semantic Planning for Video Diffusion), built on Wan2.2-T2V-A14B and paired with an MLLM semantic planner. Unified text-to-image, image editing, text-to-video, video editing, reference-to-video; 480p/16fps default, FlashAttention-3 on Hopper, video-edit quality ranked first-tier vs closed-source commercial models.

  • Luma Ray3.2

    Luma AI · Text-to-Video

    Luma's video model adding up to 16 keyframes per clip for directed transitions, Modify Video V2 (reshapes existing footage up to 20s at 1080p preserving lighting/performance), motion transfer and aspect-ratio reframing. Cinema-grade 1080p with native 16-bit HDR and EXR frame export for grading/compositing/VFX.

  • Reshoot Anything

    Adithya Iyer et al. · Video Editing

    Self-supervised diffusion-transformer model that renders new camera paths through existing video, learning 4D scene structure from monocular video with no 3D supervision by treating two crop trajectories of one clip as source/target views and using a forward-warped anchor to expose newly-visible regions. Works on live-action, animation and generative footage with only a 2D motion tracker.

  • Prometheus Productions

    Prometheus Productions · Video Editing

    AI-first creative studio offering AI filmmaking, storyboarding, music/audio, advertising, fashion and game-prototyping services blending generative visuals with traditional production, plus 3-6 hour AI masterclasses.

  • Runway Gen-4.5

    Runway · Text-to-Video

    Runway's flagship text-to-video and image-to-video model (2-10s clips) that tops the independent Video Arena ranking ahead of Veo 3 and Sora 2 Pro on realism, fluidity and physical plausibility, now available via the Runway API and app.

  • Runway Gen-4.5 — Announcement

    Runway (@runwayml) · Text-to-Video

    Runway's launch post for Gen-4.5 — see the linked X post for benchmark clips and rollout details.

  • SCAIL-2 — Project Page

    teal024 / Z.ai · Video Editing

    Project page for SCAIL-2's end-to-end character animation: conditions directly on driving-video latents (no skeleton/mask intermediates) via in-context mask conditioning + mode-specific RoPE, trained on synthetic MotionPair-60K with Bias-Aware DPO to fix hand/finger artifacts. Handles single/multi-character animation, replacement and zero-shot animal/egocentric driving.

  • StreamForce

    Northeastern University (NEU-VI) · Video Inpainting

    Streaming image-to-video model that responds to real-time user force inputs (localized pushes, global effects like wind) through a unified force representation and causal processing, reacting as forces arrive rather than needing the full sequence upfront. SOTA force adherence and motion realism at up to 16.6 FPS on a single GPU, with emergent gravity/mass/friction behavior.

  • MilliVid

    MIT & Toyota Research Institute · Text-to-Video

    Long-video generation using a hierarchical autoencoder (coarse levels capture layout/semantics, fine levels add texture) and a coarse-to-fine rollout that generates many low-detail frames then progressively refines them, keeping a constant token count (~10x more context than naive autoregressive). On Minecraft it sustains geometric consistency and object permanence far longer than FramePack at equal compute.

Audio & Speech (4)

  • Miso-TTS

    Miso Labs · Text-to-Speech

    Open-source text-to-speech foundation model for voice agents, claiming ~110ms response latency (vs ElevenLabs ~700ms) and voice cloning from a 10-second clip. Designed for on-prem/local deployment for data sovereignty, with API access and enterprise hosting.

  • ElevenLabs Music v2

    ElevenLabs · Music Generation

    Text-to-music model that transitions genres mid-track (opera → heavy metal), keeps fast/dense rap coherent, embeds non-musical SFX, and supports section-level inpainting to regenerate parts of a song. Trained only on licensed data and cleared for commercial use; shipped across ElevenMusic, the API and ElevenCreative with API pricing cut up to 50%.

  • Gemini Live 3.5 Translate

    Google · Speech-to-Text

    Real-time speech-to-speech translation that processes audio continuously (not turn-by-turn), auto-detects input language, handles noise and preserves speaker intonation/pacing/pitch while staying a few seconds behind. 70+ languages / 2000+ pairs in one meeting via the Live API, AI Studio, Google Meet and Translate, with SynthID watermarking.

  • dots.tts

    RedNote hilab · Text-to-Speech

    2B-param fully continuous, end-to-end autoregressive TTS that skips discrete tokens, pairing a semantic encoder, an LLM and an autoregressive flow-matching acoustic head over a 48kHz AudioVAE. Cross-lingual voice cloning, emotional expressiveness and context-aware generation across 24 languages (0.94% WER Chinese, 83.9 avg speaker similarity); Apache-2.0.

3D & Spatial (10)

  • Shotblock

    Shotblock · 3D Generation

    Browser-based 3D storyboard and shot-planning tool for filmmakers — block camera angles, composition and sequences in a spatial 3D environment before production.

  • Tripo Project Eden

    Tripo / VAST AI Research · 3D Reconstruction

    Persistent multiplayer world model that decouples world state from rendering via three layers: an evolving structured global state (geometry, semantics, action consequences), a state-to-observation interface, and a generative neural renderer. Supports persistent off-camera objects, multi-user state sync, editable worlds and physically consistent multi-viewpoint rendering.

  • Project Eden — Announcement

    Tripo AI (@tripoai) · 3D Reconstruction

    Tripo's launch post for Project Eden, their persistent generative world model — see the linked X post for demo clips and details.

  • WorldString

    Tsinghua IEI, UCSD, Caltech, NVIDIA · 3D Reconstruction

    Builds neural digital twins of real-world objects from sparse keypoints, learning token-based representations of 3D point clouds to reconstruct geometry and deformation over time from RGB-D capture. Demonstrated on articulated objects (robot hands), skinned characters, soft/deformable items and quadrupedal robots — actionable representations for robotic interaction.

  • World Tracing

    World Tracing authors · Depth Estimation

    Converts a single image into layered camera-space 3D geometry (six depth layers per pixel) via a world-tracing diffusion transformer treating layers as denoising tokens, using pixel-space flow matching with a mixed noise schedule. Maintains 2D-to-3D correspondence and beats both depth estimators and image-to-3D generators — enabling training-free textured meshes, 3D scene editing and geometry-guided video.

  • Flex4DHuman

    Andy Cheng et al. · 3D Reconstruction

    Method for flexible 4D (dynamic 3D-over-time) human reconstruction and generation — see the project page for video results and the full method.

  • VideoMDM

    VideoMDM authors · 3D Reconstruction

    Diffusion framework generating 3D human motion trained only with 2D video supervision: a denoiser recovers clean 3D motion while loss is applied entirely in 2D by reprojecting predictions against keypoints, with depth-aware reprojection weighting proven equivalent in expectation to 3D supervision. 0.88 FID on HumanML3D (vs 0.54 full-3D) and human-preferred on real Fit3D/NBA video.

  • SurFlo

    Antoine Guédon et al. · 3D Reconstruction

    Feed-forward 3D surface reconstruction fusing multiple unposed images into a fixed-size global latent (K=128 tokens, ViT + Perceiver), then decoding surface points independently via per-point flow-matching ODEs at any resolution, with inference-time Gaussian-splatting render guidance. Across eight benchmarks (2-32 views) it beats VGGT, Gaussian Wrapping and NOVA3R on Chamfer Distance and F1.

  • MoVerse

    Orange 3DV Team · 3D Generation

    Turns a single image into a navigable real-time 3D world in three stages: diffusion-based 360° panorama generation, lifting to a persistent 3D Gaussian-splat scaffold, then streaming photorealistic video along user camera paths. Renders at 8 FPS on a single RTX 4090 while keeping geometry and temporal coherence by separating world construction from observation rendering.

  • MeshFlow

    MeshFlow authors (CVPR 2026 Highlight) · 3D Generation

    Generates artist-quality 3D meshes with continuous geometry and explicit connectivity in ~1 second (~18x faster than autoregressive mesh decoders) by pairing a MeshVAE that compresses meshes into a continuous unquantized latent (vertex positions, normals, edge embeddings) with a flow-based diffusion transformer denoising all tokens in parallel. Unconditional, point-cloud and image-to-mesh.

Agents & Automation (4)

  • OSCAR

    Peking University, U-Michigan, NVIDIA · Embodied AI

    Omni-Embodiment Action-Conditioned world model for robotics: a 2B-param model that predicts future states from actions across diverse robot embodiments/morphologies, supporting policy evaluation and planning. Released with arXiv paper, code, HF model and training/eval datasets.

  • Arbor

    Renmin University & Microsoft Research · Agent Framework

    Generalist autonomous-research agent built on Hypothesis-Tree Refinement: a long-lived coordinator + short-lived executors grow a persistent tree linking hypotheses, artifacts, evidence and distilled insights, pruning failures and propagating lessons. Best held-out results on all six tested research tasks at >2.5x the average gain of Codex/Claude Code; 86.36% Any Medal on MLE-Bench Lite with GPT-5.5.

  • Kimi Code

    Moonshot AI · Coding Agents

    Coding-focused AI assistant built on the updated K2.7 Code model, working across terminals, IDEs and CLI for writing/debugging/refactoring, codebase analysis, shell/file automation and web research. Bundled into Kimi membership for agentic programming.

  • AnchorWorld

    AnchorWorld authors · Embodied AI

    Framework for interactive first-person world models driven by embodied human motion, using Hybrid-View Human Action Control (combining third-person training data where the full body is visible with first-person data) to project 3D motion into 2D observations, plus Evolvable Anchor-View Customization (RGB/3D-pose/text priors). Lets egocentric worlds evolve while preserving spatio-temporal geometry.

Developer Tools (6)

  • comfy-cli

    Comfy-Org · Inference Engine

    Python 3.10+ CLI for installing, launching and managing ComfyUI: one-command install, custom-node install/update/snapshot with bisect + unified dependency conflict detection, model downloads from CivitAI/HuggingFace, and direct API calls to hosted Flux/Runway/Kling/DALL-E. Runs workflows against local servers with auto UI-to-API JSON conversion.

  • train-llm-from-scratch

    Fareed Khan · AI Infrastructure

    Educational PyTorch repo implementing a decoder-only transformer from scratch (token+position embeddings, multi-head attention, stacked blocks), scaling 13M → 2.1B params (64 blocks) on a subset of The Pile with tiktoken + AdamW. Includes a full post-training suite: SFT, reward modeling, PPO, DPO, GRPO.

  • CLI Printing Press

    Matt Van Horn · Web Scraping

    Generator that turns OpenAPI specs, HAR files, or live website URLs (reverse-engineered via browser automation) into agent-optimized Go/Cobra CLIs plus MCP servers, each backed by a SQLite store with full-text search. Agent-native output: auto-JSON when piped, --compact mode cutting tokens 60-80%, typed exit codes, 4-stage verification before deploy.

  • Get It

    beltromatti (GDG AI Hack Milan) · AI Infrastructure

    Electron/Next.js + React 19 study app that parses text PDFs (pdf.js), auto-detects concepts and inserts inline Three.js/Canvas visualizations (3D models, formulas, graphs), and tracks mastery via a knowledge graph scored on memory/comprehension/structure/application. Study modes: scoped chat, flashcards, quizzes, Feynman sessions — all local on the user's own OpenAI key.

  • PriceGhost

    clucraft · Web Scraping

    Self-hosted TypeScript price-tracker (React 18 + Node/Express + PostgreSQL + Puppeteer) monitoring any e-commerce product via multi-strategy extraction: JSON-LD parsing, site-specific scrapers, generic CSS selectors, and AI fallback (Claude/GPT/Ollama) that vote on the correct price. Sends drop/back-in-stock alerts via Telegram, Discord, Pushover, ntfy, Gotify.

  • MiniMax Sparse Attention (MSA)

    MiniMax AI · AI Infrastructure

    CUDA kernel library for NVIDIA SM100 GPUs shipping dense FlashAttention plus sparse top-k block attention via two JIT stacks (csrc + CuTe-DSL), with paged FP8 decode and BF16/FP8/NVFP4/FP4 quantization. A cheap proxy pass selects high-scoring attention blocks, then the sparse kernel computes only those — the attention powering MiniMax M3's 1M-context speedups.

Research & Safety (4)

  • jennykrakovsky — Shared Post

    Jenny Krakovsky (@jennykrakovsky) · Survey & Overview

    Shared AI product/demo post — see the linked X post for the prompt and full details.

  • ai_gezgini — Shared Post

    @ai_gezgini · Survey & Overview

    AI-news/curation post highlighting a notable product or demo — see the linked X post for full context and any prompt.

  • YouTube AI Labels Update

    YouTube · Survey & Overview

    YouTube makes AI-disclosure labels more prominent (below the player on long-form, overlaid on Shorts) and adds automatic detection that labels significant photorealistic AI generation, with creator override in Studio. Labels are permanent for Veo/Dream Screen content or C2PA fully-generative metadata; a label alone doesn't affect recommendations or monetization.

  • Agents' Last Exam

    UC Berkeley RDI · Benchmark

    AI-agent benchmark measuring performance on economically valuable, long-horizon professional computer tasks with objectively verifiable outcomes across 55 sub-industries (After Effects VFX, Siemens NX 3D modeling, Unreal game dev, Rhino architecture, FSLeyes brain imaging). 1,500+ tasks toward a 5,000 target, built with 300+ industry experts.

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