Quanta AI Labs

33 / Aug 10 – Aug 16, 2026 · 56 RELEASES

AI News Week 33, 2026

Every Lab Shipped At Once + A Crack in Encrypted Reasoning. Six frontier releases inside seven days: DeepSeek-V4-Pro at 1.7T with a 1M context, Qwen3.8's 2.4T open weights actually landing as promised, Grok 4.6, Gemini 3.7 Flash posting a 16-point DeepSWE jump three weeks after 3.6, GLM-5.3 taking CyberGym on post-training alone, and Meta's 30B Muse Glimmer built for local agents rather than chat. The quieter story is speed and size collapsing in opposite directions - OpenAI's Ultrafast runs GPT-5.6 Sol at 750 tokens per second on Cerebras, LTX-2.5 makes ten seconds of 720p video in 6.8, and Cactus fits a working tool-calling agent into 45M parameters and a 14MB binary that decodes on a Raspberry Pi. The week's most uncomfortable paper has nothing to do with capability: encrypted reasoning traces turn out to be interchangeable across sessions, users and models, and 315,320 harvested blocks gave up 367 PII artefacts and 182 credentials across Anthropic, OpenAI and Google. From India, Sarvam and AI4Bharat opened Indic DiarBench across all 22 scheduled languages, 485 speakers, 189 districts.

Language Models (9)

  • DeepSeek-V4-Pro-0813

    DeepSeek AI · MoE Architecture

    A 1.7-trillion-parameter MoE with a DSpark speculative-decoding module and a 1M-token context, released MIT. Posts 87.9 on Terminal-Bench 2.1, 62.7 on DeepSWE, 74.1 on Toolathlon-Verified and 60.0 on HLE with tools. Exposes three reasoning-effort levels and will emit up to 384K output tokens in the higher modes.

  • Qwen3.8-2.4T-A95B (open weights)

    Alibaba Qwen · MoE Architecture

    The open weights promised when Qwen3.8-Max launched last week. 2.4T total, 95B active, 92 layers built as 23 cycles of three Gated DeltaNet MoE blocks to one Gated Attention MoE block. Context is 262,144 native, extensible to 1,010,000. Scores 92.6 on GPQA Diamond, 86.6 on Terminal-Bench 2.1 and 67.7% on SWE-bench Pro. Thinking mode is mandatory.

  • Qwen3.8-27B

    Alibaba Qwen · Reasoning

    The dense sibling: 27B with a vision encoder and the same hybrid Gated DeltaNet plus Gated Attention stack over 64 layers, Apache-2.0 and a 262,144-token context reaching 1M. Hits 90.3 on LiveCodeBench v6, 89.2 on GPQA Diamond, 84.3 on OSWorld computer use and 81.9 on AndroidWorld. Reasoning depth is a per-request parameter.

  • Grok 4.6

    xAI · Reasoning

    Released 12 August with the stated focus on long-running agents and turning a broad product idea into a working first version. Ties GPT-5.6 Sol at 61 on the AA Intelligence Index and leads GDPVal-AA v2 at 1753, but trails badly on DeepSWE v1.1 at 65.9% against Sol 73%. Priced $2 in / $6 out per million; the fast variant doubles that.

  • Gemini 3.7 Flash

    Google · Reasoning

    Shipped 13 August, three weeks after 3.6 Flash, and the jumps are large for a point release: DeepSWE v1.1 goes 49.0% to 65.3%, FrontierCode 1.1 34.4% to 43.6%, AutomationBench 17.0% to 30.4%. Introductory pricing of $0.75/$3.75 per million holds until 31 December 2026, then doubles to $1.50/$7.50.

  • GLM-5.3

    Z.ai · MoE Architecture

    Launched 14 August on the same 743B base as GLM-5.2 - roughly 40B active - with every gain coming from further RL and training-environment design rather than a new pre-train. Takes CyberGym at 84.5%, ahead of Mythos 5 (83.8) and GPT-5.6 Sol (83.6), and reports 2,436 confirmed vulnerabilities across 269 open-source projects. Weights staged for late August.

  • Muse Glimmer

    Meta · Reasoning

    A 30B Apache-2.0 model built specifically for local agents rather than chat: reliable tool-calling, state that persists across restarts, and self-managed memory. Reports 76.0 on SWE-Bench Verified, 75.5 on MCP Atlas and 94.7 on AIME 2026, with multimodal perception, and is sized to run on a single GPU or consumer hardware.

  • Nemotron 3.5 Lightning & NeMo Switchyard

    NVIDIA · MoE Architecture

    An 11 August pair: a 30B MoE tuned for specialist roles inside multi-agent systems, claiming up to 4x faster output and 30% faster agentic task completion, plus Switchyard, an open-source router that sends each request to the cheapest model that can still do it - NVIDIA reports frontier-level accuracy at roughly a third the cost of Opus 4.8 alone.

  • Laper

    Laper Corp · Text Generation

    Screenwriting software built by a former screenwriter that deliberately diagnoses rather than generates: an AI script doctor flags structure, character arc, pacing and motivation problems while the writing stays yours. Adds outline and card views, real-time collaboration over Loro CRDT, and generated storyboard frames. Built with Claude Code.

Vision & Image (2)

  • Sign Language to Text (SL2T)

    Google DeepMind · Visual Understanding

    Launched 12 August on Pixel 11 at no cost: on-device MediaPipe pose tracking feeds a translation model that maps coordinate sequences straight to text, skipping the intermediate gloss annotation the field has leaned on. Scores 70 BLEURT zero-shot on FLEURS-ASL, well above any previously reported figure. Trained on 100,000+ hours across 50+ sign languages.

  • UniFace

    yakhyo · Visual Understanding

    One Python API across 15 face-analysis tasks that normally means stitching five libraries together: detection via six architectures including RetinaFace, SCRFD and BlazeFace, recognition via AdaFace, ArcFace and EdgeFace, 106/98/68-point and 3D-mesh landmarks, gaze, parsing, matting, tracking with persistent IDs and anti-spoofing. CPU, Apple Silicon or CUDA. MIT.

Video & Animation (7)

  • LTX-2.5

    Lightricks · Text-to-Video

    A 22B open-weights model released 11 August that produces synchronised video and audio in one pass. Generates a 10-second 720p image-to-video clip in 6.8 seconds on two GB200s; the managed API took 23.7 seconds at 1080p. Adds native multishot scenes holding identity and lighting across cuts, 4K HDR and RAW output, and a Gemma 4 12B text encoder.

  • The Foundation Film Is Made On

    Lightricks · Text-to-Video

    The positioning piece behind LTX-2.5: Lightricks argues the useful unit is no longer the clip but the shot sequence, and that a model holding character, environment, lighting and voice across cuts is what makes generated footage editable rather than disposable. Commercial use stays free under the LTX-2.x Community License for companies below $10M revenue.

  • MAGI-2-preview

    Sand AI · Text-to-Video

    A 114B MoE with roughly 6B active per token that generates 10-second clips with synchronised audio from text or from text plus a still. Runs a two-stage pipeline - a low-resolution preview pass then a refiner to 1080p at 1088x1920 - and ships Apache-2.0. Needs eight NVIDIA Hopper GPUs to run, which places it firmly in lab territory.

  • JoyAI-Video-Edit

    JD.com · Real-Time Video

    Edits a live camera stream from natural-language instructions without waiting for the whole video, pairing an MLLM condition encoder with a causal video VAE and a multimodal diffusion transformer under bounded KV-state inference. Hits 30 FPS end-to-end at 720x1248. Apache-2.0, 1.4k stars; code landed 5 August with upgraded checkpoints on 14-15 August.

  • SCoPE

    HKU, HKUST & Tencent ARC · Text-to-Video

    Gives every token its camera ray as a second positional coordinate, making camera control part of the coordinate system instead of a bolted-on control module. Retrofits onto Wan2.2 for under 0.1% new parameters via a normalise-gate-inject step that preserves RoPE. Cuts rotation error 29% and improves FVD 43% at 14B scale. Code and weights are out.

  • Diffusion Studio Editor

    Diffusion Studio (YC F24) · Video Editing

    A desktop editor plus a CLI (dapi) built so agents can do professional video editing rather than just generate clips: analyse media, build compositions in JSX over SolidJS, cut, add motion graphics and titles, then hand a human the same timeline to refine. Includes transcription, waveform analysis and frame extraction. MPL-2.0, 499 stars.

  • YouTube Clipper Skill

    op7418 · Video Editing

    Claude skill that pulls a YouTube video, reads it to generate semantic chapters of two to five minutes, clips the segments with FFmpeg, translates subtitles into a bilingual track and optionally burns them in. Batches the translation to cut API calls by a claimed 95%, and formats output for Xiaohongshu and Douyin. MIT, 2.1k stars.

Audio & Speech (3)

  • Indic DiarBench

    Sarvam AI & AI4Bharat · Speech-to-Text

    An open joint diarisation and ASR benchmark covering all 22 scheduled Indian languages: ~108 hours of natural multi-speaker audio, 485 speakers drawn from 189 districts, two to nine speakers per recording with real overlap, code-mixing and interruption. Human-corrected speaker-attributed transcripts throughout. Sarvam own pipeline leads at 16.0% DER and 38.8% cpWER.

  • MiniMax Music 3

    MiniMax AI · Music Generation

    Generates complete songs up to five minutes at 32kHz stereo from two inputs: a music caption covering style, mood and arrangement, and lyrics tagged by section. Uses a hierarchical autoregressive stack - an 8B global model for long-range structure, a 0.6B local model for frame detail, and Flow Matching with a Flow-VAE for synthesis. Native ComfyUI template.

  • MiDashengLM-Gen

    Xiaomi · Audio Processing

    Generates whole audio scenes - speech, music, effects and ambience mixed together - from a text description at 16kHz, rather than treating speech synthesis as a separate problem. Runs a fully fine-tuned Qwen3-1.7B backbone over a DashengTokenizer producing 768-dimensional latents at 25Hz, with a 16-layer Flow Matching DiT and a learned stop head. Apache-2.0.

3D & Spatial (2)

  • Hunyuan3D-WorldClaw

    Tencent Hunyuan · 3D Generation

    Builds an editable 3D open world from one text prompt by orchestrating Claude Opus 4.8, GPT-Image-2, SAM3D and Hunyuan3D inside Blender - not by training one giant world model. Terrain is procedural code the agent writes and parameterises, and a render-inspect-refine loop catches scale and placement errors. Output is independent textured meshes. No code or weights yet.

  • Unreal Game Assets Creation Skill

    LaurentiuGabriel · 3D Generation

    Claude Code skill that turns a prompt or a photo into a game-ready 3D asset by driving four local services headlessly in sequence: Fooocus for the image, Hunyuan3D-2 for the mesh, Blender for format conversion, then optional import into Unreal Engine. Textures are embedded in the output. Runs on one Windows workstation with 8GB of VRAM. 117 stars.

Agents & Automation (10)

  • Pi Agent Toolkit

    earendil-works · Agent Framework

    Moved from Week 23. Now 91.1k stars across 5,685 commits. Recent work is unglamorous and telling: supply-chain hardening with pinned dependencies, lockfile integrity checks and scheduled security audits, plus documented containerisation patterns for sandboxing the agent. New contributors issues and PRs are auto-closed by default. MIT.

  • Agency Agents

    msitarzewski · Agent Framework

    Moved from Week 11. Now at roughly 146,000 stars and 23,500 forks, and it has grown a native desktop app for macOS, Linux and Windows installable by homebrew cask. The roster is 230+ agents across 18 divisions - engineering alone has 60+ - with installers for 14+ coding tools including Claude Code, Copilot, Cursor and Windsurf. MIT.

  • ML Intern

    Hugging Face · Agent Framework

    Moved from Week 17. Hugging Face's autonomous ML engineer sits at 10.7k stars and has added local-model support through any OpenAI-compatible endpoint (Ollama, vLLM, LM Studio), GPU testing in a sandboxed HF Space, Slack approval gateways, doom-loop detection for repeated tool failures, and context auto-compaction at 170k tokens. Apache-2.0.

  • DeepSeek Harness (dsh)

    DeepSeek AI · Agent Framework

    DeepSeek own open agent harness, plugin-based on Cordis with a web UI on port 3080, sitting at roughly 123.8k stars and 12,293 commits. Still labelled developer preview, and the README says plainly that there will be compatibility-breaking changes - worth weighing before building on it. MIT licensed.

  • Cloudflare OS

    Cloudflare · Agent Framework

    An agent platform on Workers where users ask the AI to build small sandboxed apps (Gadgets) and share them as reusable Blueprints. The interesting part is Gatekeepers: capability-based access control sitting between agents and external services, with human approval in the loop. Built on Durable Objects and the open-source workerd runtime. Apache-2.0, 8.4k stars.

  • LifeOS

    Daniel Miessler · Agent Framework

    Sits on top of Claude Code or Cursor as a personal context layer: a persistent memory store (Cortex), an Ideal State Artifact for goal tracking, a seven-phase problem-solving loop, and a Pulse daemon giving it voice, hooks and a dashboard. Routes tasks by context rather than treating every request the same. MIT, 18.5k stars, currently v7.40.4.

  • AgentCall

    sidinsearch · Voice Agents

    Turns a rooted Android phone into a voice endpoint for a local agent over MCP, so Hermes or OpenClaw can place and take real cellular calls on your SIM. Speech runs through OpenAI, ElevenLabs or local Supertonic; contacts, call history and recordings stay on the handset. Fails closed on policy or health errors and only talks to the desktop over authenticated USB. AGPL-3.0.

  • Vobiz AI

    Ilailimitado · Voice Agents

    Carrier-grade voice infrastructure aimed at AI applications and contact centres: SIP trunking, voice APIs, number provisioning, WebRTC and WhatsApp, pay-as-you-go with no setup fee. US local outbound runs $0.0115/min and inbound $0.0055/min, numbers $1.00/month, speech recognition $0.0205 per 15 seconds, transcription $0.0098/min.

  • TradingAgent

    marketcalls · Agent Framework

    Chat-first trading for Indian markets wiring Agno to OpenAlgo, and the design is unusually careful: every order stops for explicit human approval, analyzer mode simulates by default, RiskGuard runs deterministic post-approval checks, orders are logged twice, and a KILL file halts trading without a restart. 127 indicators available. MIT.

  • Dyna-2

    Dyna Robotics · Robotics

    A world-action model trained on over a million hours of human egocentric video that predicts future frames and robot actions together, then transfers zero-shot to robot bodies it never saw. Accuracy@0.5 climbs 0.067 to 0.159 as data scales 1k to 1M hours, and on-robot mean score goes 20% to 53%. Bottle-cap untwisting succeeds half the time on ten minutes of robot data.

Developer Tools (20)

  • Ultrafast mode (GPT-5.6 Sol)

    OpenAI · Inference Engine

    A latency tier, not a new model: the existing GPT-5.6 Sol running on Cerebras hardware at up to 750 output tokens per second, up to 14x faster than standard, with intelligence claimed unchanged. In limited preview with selected customers testing it on coding, commerce, financial research and support. No pricing and no general date announced.

  • OmniRoute

    Diego Souza · AI Infrastructure

    An AI gateway putting 340+ providers behind one OpenAI-compatible endpoint with 19 routing strategies and automatic fallback across subscription, key, budget and free tiers. Stacks 12 token-compression engines claiming 15-95% savings, and documents roughly 1.51 billion free tokens a month across deduplicated free tiers. Runs local-first with no telemetry. MIT, 48.8k stars.

  • Repowise

    repowise-dev · AI Infrastructure

    Indexes a repo into a dependency graph across 19 languages, generated architecture docs and a code-health score from 49 deterministic detectors, then exposes ten task-shaped MCP tools to agents. Reports 0.876 file coverage against CodeGraph 0.610, and 31.6% fewer output tokens with 3.8 tool calls versus 7.2 for a bare agent. AGPL-3.0, 5.9k stars.

  • BitNet

    Microsoft · Inference Engine

    Microsoft's inference framework for 1-bit (ternary) LLMs reports 2.37x to 6.17x speedups on x86 with 71.9-82.2% less energy, and 1.37x to 5.07x on ARM. July added VibeASR.cpp for multilingual speech recognition under BitNet quantisation plus 0.6B and 270M embedding models. MIT, 40.1k stars - and still no tagged GitHub release.

  • h3.c

    antirez · Inference Engine

    Salvatore Sanfilippo native MiniMax-H3 inference engine for Apple Silicon, in C with Metal acceleration. Renders 512x512 at 20 steps in about 16.69 seconds on an M5 Max, 12.60 with token reduction, and roughly 3.5 seconds in aggressive four-step mode. Produces 22-frame clips at 24fps, supports first/last-frame anchoring and ordered reference conditioning. MIT.

  • MiniMax-H3-FineTuning

    IAmIronMan42 · AI Infrastructure

    MiniMax shipped H3 weights but no trainer; this fills the gap with a two-stage supervised framework - offline latent caching of video, audio and text, then transformer optimisation under rectified-flow loss. Trains heads (~1M params), LoRA (~10.7M) or the full 33B, with DeepSpeed ZeRO-3 and a ~70k-token sequence ceiling. Practical runs want 8x A800-80GB. 304 stars.

  • NVIDIA Nsight AI

    NVIDIA · AI Infrastructure

    The rebrand of Nsight Copilot into a platform for CUDA development: an authenticated CUDA MCP Server pointing agents at current NVIDIA docs and examples, a self-hosted open-source Copilot Blueprint, AI-guided analysis inside Nsight Compute, and a VS Code extension for CUDA-aware chat and code transformation. Gives concrete guidance on issues like uncoalesced memory access.

  • here.now

    here.now · AI Infrastructure

    Hosting built for agents rather than people: an agent turns a file, folder, report or dashboard into a live URL at slug.here.now, anonymously for 24 hours or permanently with an API key. Adds version history with restore, password or invite-only access, analytics, and a Site Data store letting static sites save validated records without a backend.

  • AgentSearch

    brcrusoe72 · Web Scraping

    Self-hosted search for agents: SearXNG across 25 engines wrapped in a FastAPI layer of 17 endpoints that deduplicates across engines, scores results and escalates extraction from direct fetch through readability, browser render, Wayback and PDF handlers. Scrubs prompt injection, detects paywalls, and analyses its own fetch failures. No third-party API keys. MIT.

  • Skill Recorder

    Microsoft · Prompt Engineering

    Records you doing a task on screen, then uses the GitHub Copilot CLI to reconstruct it as an intent statement plus ordered steps and emit a reusable Skill or Automation for Microsoft Scout, Copilot Cowork or Copilot Studio. Transcription runs on-device across 99 languages. macOS primary, Windows 11 supported. MIT, 3.2k stars.

  • Diagram Design

    Cathryn Lavery · Prompt Engineering

    Claude Code skill generating 29 diagram types as self-contained HTML and SVG that open offline with no build step - explicitly positioned against Mermaid output. Extracts your colours and fonts from your website in a 60-second onboarding, imports draw.io and Mermaid for redrawing, and ships ARIA labels and WCAG AA contrast by default. MIT, 19.1k stars.

  • Oil Motion

    oil-oil · Prompt Engineering

    Agent skill that turns AI-generated video sequences into interactive web animations driven by scroll, mouse, drag, touch or device orientation. Strips pauses and duplicate frames, optimises assets separately for desktop and mobile, validates frame consistency and motion smoothness, and outputs either alpha WebP image sets or green-screen MP4. MIT, 1.8k stars.

  • LTX-2.5 Agent Prompt Skill

    Benji AI Playground · Prompt Engineering

    Turns a rough video idea plus optional reference image or audio into an LTX-2.5 prompt written as flowing cinematic prose rather than JSON. Offers single-shot, multi-shot (two to four connected cuts in one prompt) and screenplay modes over a six-part structure covering shot type, lighting, action, character, camera and audio, with 100+ categorised cinematic terms. MIT.

  • SuperBrain

    sidinsearch · Data Integration

    Self-hosted personal archive: share a link from Instagram, YouTube or a browser into an Android app and a Python backend summarises, categorises and tags it into a local SQLite database you own. Identifies background music in reels via Shazam, transcribes audio through Groq Whisper with local Whisper as offline fallback, and routes across Groq, Gemini, OpenRouter and Ollama. AGPL-3.0.

  • MatrAIx

    MatrAIx · Evaluation Benchmarks

    Evaluation infrastructure that tests products against simulated users before real ones see them: a population of 8.3 billion persona agents carrying 1,290+ attributes, run across surveys, chatbot testing, web prototype evaluation and app testing. Returns per-segment reports, telemetry and interaction trajectories rather than a single aggregate score.

  • Daily Stock Analysis

    ZhuLinsen · BI & Analytics

    Generates a daily decision dashboard across A-shares, Hong Kong, US, Japan, Korea and Taiwan by combining market data, live news and technical indicators, then pushes it to WeChat Work, Feishu, Telegram, Discord, Slack or email. Ships 15+ built-in strategies queryable by agent, and runs free on GitHub Actions with no infrastructure. MIT, 63k stars.

  • Deltalytix

    Hugo Demenez · BI & Analytics

    Open-source trading journal for futures and prop-firm traders that syncs directly from Rithmic, Tradovate, Thor and DXfeed, imports from Interactive Brokers, ATAS and FTMO with AI-assisted CSV mapping, and runs an AI journal that names the patterns costing you money. Team accounts aggregate prop-firm stats. CC BY-NC 4.0, self-hostable.

  • Journedge

    TheQuantum-Dev · BI & Analytics

    Desktop trading journal that keeps everything local in SQLite. Imports from Fidelity, Schwab, TD Ameritrade, Tastytrade and Interactive Brokers, then computes Sharpe, Sortino and Calmar alongside hour-of-day heatmaps and behavioural analysis. Adds pre-trade planning so you can compare plan against execution, plus daily loss and trade-count limits. MIT, v4.0.0.

  • BharatRajya

    The Kaipullai · BI & Analytics

    An interactive district-level historical atlas of the Indian subcontinent: drag a year slider and watch who ruled each district change hands, with kingdoms, kings and battles layered in and a state-wise view for tracing one region across centuries. Built largely by a non-coder directing AI tooling, and picked up publicly by Vijay Shekhar Sharma among others.

  • BharatRajya - 1188 CE

    The Kaipullai · BI & Analytics

    The deep end of the same atlas. 1188 CE currently marks the edge of its coverage - the maker has said publicly that extending data beyond that year is the active work - which makes this page a useful check on how thin the sourcing gets the further back an AI-assisted history project reaches. Worth reading as a data-provenance question, not just a map.

Research & Safety (3)

  • Sol-Attn

    Li, Ye, Liu, Xie, Han et al. · Research Paper

    Training-free sparse attention for video diffusion transformers that unifies dynamic routing and sparse computation through on-the-fly block thresholding with proxy-score reuse, so the full proxy score map never has to be materialised. Delivers 2.1x end-to-end speedup on video generation and 2.3x on editing while holding visual quality. Submitted 27 July.

  • Stealing Reasoning Traces from Proprietary LLM APIs

    Panfilov, Shumailov, Geiping, Andriushchenko et al. · Red Teaming

    The encrypted reasoning blocks the major labs return turn out to be interchangeable across sessions, users and models - so injecting them into a weaker model decrypts a stronger one private reasoning without jailbreaking it. From 315,320 blocks decoded out of public repositories the authors recovered 367 PII artefacts and 182 credentials. Demonstrated against Anthropic, OpenAI and Google.

  • Awesome AI in Finance

    georgezouq · Survey / Overview

    Long-running curated index of LLMs, deep-learning strategies and tooling for financial markets, now spanning trading agents and multi-agent systems, MCP servers for market data and execution, academic papers, backtesting frameworks, exchange APIs and data sources across equities, crypto and prediction markets. CC0-1.0, 6.4k stars, 259 commits.

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