Quanta AI Labs

32 / Aug 3 – Aug 9, 2026 · 29 RELEASES

AI News Week 32, 2026

Trillion-Parameter Week + Agents Grow a Memory. Alibaba's Qwen3.8-Max lands at 2.4T parameters with 95B active and a 1M context, taking Terminal-Bench 2.1 to 86.6 and IFBench to 82.8, with open weights promised for the following week - while Meta finally enters the coding-agent fight with Muse Spark 1.2 and Muse Code, though independent Vals testing puts it 14th on the common harness against its own 82.9% claim. The counter-story is compression: kimi-k3-in-c runs a 2.78T-parameter MoE on one CPU in 8.24GB of RAM in plain C99, and BigBang-v1 claims best-in-class among 35B models on eight benchmarks with only 3B active. Agent tooling converged on the same insight from three directions - LongHorizon-Harness, Prime Agent and Graft all attack context rot by refusing to let unverified work accumulate. Elsewhere: MiniMax-H3 generates video with native stereo audio in one pass, DeepMind's WeatherNext Cyclones adds a day of hurricane lead time in Nature.

Language Models (2)

  • Qwen3.8-Max

    Alibaba Qwen · MoE Architecture

    A 2.4T-parameter MoE with 95B active per query and a 1M-token context, natively accepting text, images and video. Scores 86.6 on Terminal-Bench 2.1 (above Claude Opus 4.8 and Fable 5 at 84.6, below GPT-5.6 Sol at 88.8), leads PaperBench at 93.0 and IFBench at 82.8. Priced $2 in / $6 out per million tokens; open weights due the week of 10 August.

  • BigBang-v1

    Endless Frontier · MoE Architecture

    Agentic model at 35B total and 3B active, fine-tuned from Qwen3.6-35B-A3B, operating a tool loop of web search, page reading and code execution. Evaluated on BrowseComp, xbench, SWE-Bench Pro, SciCode-Verified, FrontierScience-Research and Humanity’s Last Exam, where it claims the highest reported score among selected 35B models on all eight. Apache-2.0 harness, 42 stars.

Vision & Image (2)

  • Boogu-Image-0.1

    Boogu · Text-to-Image

    Moved from Week 25. Huawei HK's unified open-source image generation and editing family now spans Base (50-step), Turbo (3-4 step distilled), fp8, Edit and Edit-Turbo variants. Tops open-source on Qwen-Image-Bench at 53.58 with 10B params and leads ImgEdit_O at 4.64, on training data roughly an order of magnitude smaller than rivals. Apache-2.0.

  • InvokeAI

    Invoke · Image Editing

    Long-running open-source generation and editing suite built around a Unified Canvas and a node-based workflow graph, with board-based gallery management and SAM/SAM2 object segmentation. Supports Stable Diffusion 1.5 through 3.5, SDXL, Flux variants, CogView 4 and Ideogram 4. Apache-2.0, 27.8k stars, 19,111 commits on main.

Video & Animation (5)

  • MiniMax-H3

    MiniMax AI · Text-to-Video

    A 33B dense single-stream H3-Omni-Transformer with modality-specific AdaLN branches (~13B) and 3D multimodal RoPE, generating 4-15 second video at 24fps with native 32kHz stereo audio rather than a separate audio pass. Reaches 2K through an H3-Regenerate-2K module and handles text-to-video, image-to-video and reference-to-video. MiniMax H3 Community License.

  • capcut-cli

    renezander030 · Video Editing

    Edits CapCut and JianYing projects by rewriting their local JSON directly - no API, no server, no dependencies. Handles Whisper captions, SRT/VTT import-export, translation, scene detection, chroma key, templates, FFmpeg rendering and long-form-to-shorts cuts, and runs as CLI, library or JSONL queue. Node 18+, MIT, 304 stars; v0.17.1 carries seven security fixes.

  • Muse Minimax Director (Seed Hunt)

    Muse Collective · Text-to-Video

    ComfyUI node giving MiniMax H3 a timeline: scripts split into CUTs with character and location reference images, auto-chunked to H3’s per-call ceiling of ~15 seconds with reference-tag numbering handled for you. Offers a 9-reference mode or hard-locked first/last frames, multi-speaker auto-tagging, and a Seed Hunt toggle producing four same-settings variants. MIT, v1.1.0.

  • Wan-Animate-2

    Tongyi Lab, Alibaba · Real-Time Video

    Animates a character from one reference image plus a driving video, dropping the intermediate motion extractor in favour of a dual-branch DiT with Time-Align RoPE and Sparse-Ref Attention. Handles multi-character and multi-camera scenes and decouples viewpoint control into a text prompt. A distilled Lite variant targets real-time avatars. Base weights promised; arXiv 2608.06009.

  • LeapTalk

    Harbin Institute of Technology & Shanghai Jiao Tong University · Real-Time Video

    Reframes talking-head generation as reference-anchored transformation via a Brownian bridge instead of noise-to-data diffusion, collapsing it to a single step. Heterogeneous distillation with SNR-aligned time transformation moves the teacher into the student. Hits up to 200 FPS on one H200 at one inference step per chunk, streaming arbitrarily long video. arXiv 2608.00079.

Audio & Speech (2)

  • SymphonyGen

    Central Conservatory of Music, XJTU, USTC & Shenzhen University · Music Generation

    Generates orchestral music by writing a harmony skeleton first, then orchestrating on top of it - so existing pieces can be re-orchestrated, not just composed from scratch. Uses a 3D hierarchical framework with cascading decoders over bar, track and event axes, RL refined against CLaMP 3 audio embeddings, and dissonance-averse sampling. Beats SymphonyNet and NotaGen-X.

  • VocalRender

    pymaster17 · Music Generation

    Singing synthesis straight from composer-facing notation - lyrics, MIDI pitches, durations and tempo - with no phoneme-level timing and no reference audio required. Pairs an AudioVAE for continuous acoustic latents with autoregressive diffusion for prosody, and keeps a score-native representation so note-to-word alignment and melisma survive. Code and weights on GitHub and HF.

3D & Spatial (2)

  • MAC (Multi-Agent CAD)

    Tsinghua IEI Lab · 3D Generation

    Turns text into printable STEP/STL geometry through four LLM agents on a LangGraph state machine: requirement parsing, geometric planning, code generation, then autonomous QA and repair. Every intermediate artefact stays human-auditable. On a 141-feature suite it cuts tokens 103.9M to 896K (116x) and API calls 1,307 to 50, while pass rate rises 97.9% to 99.3%. MIT, 495 stars.

  • Hunyuan3D-Buffalo 1.0

    Tencent Hunyuan3D · 3D Generation

    Tencent unifies 3D understanding, text-to-3D, instruction-guided editing and text-grounded part generation in one architecture: Hunyuan3D-VLM supplies semantic conditioning, Hunyuan3D DiT does synthesis. Trained on an 87M corpus (25M understanding, 50M text-to-3D, 12M editing pairs built with Nano3D-v2). Cuts Chamfer Distance 86.7% on Edit3D-Bench.

Agents & Automation (7)

  • minimal-embodiment

    Olivia Zhu · Embodied AI

    Gives an LLM a small physical body on an ESP32: four sensor modules (environment, light, motion, sound) and three outputs (haptic, OLED face, piezo buzzer). Two of the outputs feed back into inputs, so the model hears itself speak and feels itself tap. Ships a ~1,100-line Arduino sketch, a dependency-free Node bridge and a 660-trial validation set. MIT, 244 stars.

  • GenOffice

    Genspark AI (Mainfunc) · Desktop Agents

    Six Electron apps - Docs, Sheets, Slides, PDF, Markdown and a suite shell - over a shared TypeScript engine, with a Rust sidecar for spreadsheet maths. Claims byte-preserving round-trips for .docx/.xlsx/.pptx/.pdf and puts the AI agent inside the editing flow with version snapshots and diffs rather than beside it in a chat panel. Apache-2.0, 2.3k stars, v0.5.83.

  • Claw Code

    ultraworkers · Coding Agents

    A Rust CLI agent harness its authors describe as an agent-managed museum exhibit rather than a product: agents handle planning, execution, verification and maintenance, and the README steers anyone wanting real work toward LazyCodex or Gajae-Code instead. Sits at roughly 195k stars and 109k forks despite disclaiming production use. MIT.

  • Prime Agent

    Prime Intellect · Agent Framework

    Coding and research agent built for long-running autonomous work on two abstractions: a Recursive Language Model and a Continual Harness holding prompts, memories and reusable specs. Treats context as variables and tools as function calls inside a persistent IPython REPL, spawns subagents via rlm(), and keeps sessions alive in a daemon for reattachment. MIT, 9.2k stars.

  • Xiaomi-Robotics-1

    Xiaomi Robotics · Robotics

    Vision-language-action foundation model for mobile manipulation, pairing Qwen3-VL with a diffusion transformer at 5B parameters and trained on over 100,000 hours of real-world trajectories. Posts 74.5% on RoboCasa, 59.1% on VLABench and 57.4% on RoboCasa365. Embodiment-free pre-training then cross-embodiment post-training means hours of demos adapt it. Apache-2.0; code landed 3 August.

  • Muse Spark 1.2 & Muse Code

    Meta · Coding Agents

    Meta enters the coding-agent fight with a coding model plus a terminal agent running persistent async background agents. Muse Spark 1.2 scores 82.9% on Terminal-Bench 2.1 - behind Claude Opus 5 at 86.7%, ahead of GPT-5.6 Terra at 81.8% - and 59.3% on DeepSWE 1.1. Numbers come from Meta’s own harness; independent Vals testing places it 14th on the common harness.

  • LongHorizon-Harness

    DreamX Team · Agent Orchestration

    Replaces the ever-growing agent session with a Manage-Execute-Audit loop: a manager decomposes work from verified facts, an executor runs each subtask on fresh context, and an auditor independently certifies what actually changed before anything enters the record. On one Qwen 3.7-Plus backbone, WeaveBench pass rate goes 51.8% to 80.7% and Terminal-Bench 2.1 69.7% to 77.2%.

Developer Tools (8)

  • OpenMed

    Maziyar Panahi · AI Infrastructure

    Moved from Week 24. v2.0.0 turns the local-first clinical NER and PII de-identification toolkit cross-platform: native Swift (OpenMedKit), Kotlin/Android and Transformers.js WebGPU alongside Python, plus MLX acceleration on Apple Silicon. Adds 35 multilingual PII routes and up to 3.3x batch throughput. 2,200+ models, 21+ languages, Apache-2.0, 5,000+ stars.

  • Claude Trading Skills

    tradermonty · Prompt Engineering

    Collection of 60+ Claude Skills for equity traders covering market-regime analysis, portfolio review, swing screening, trade planning and journaling. Explicitly refuses to automate buy/sell decisions, positioning itself as structure for review and risk management instead. Optional FMP, FINVIZ Elite and Alpaca integrations, plus a no-API path on public CSVs. MIT, 2.6k stars.

  • Trading Skills

    staskh · Data Integration

    Options-trading advisor exposing 32 MCP tools and 25 skills to Claude Code, Cursor or Claude Desktop: live quotes, option chains, RSI/MACD/Bollinger, Greeks and spread analysis, PMCC scanners, earnings calendar and Interactive Brokers portfolio access. Yahoo Finance free tier runs 15 minutes delayed. Python 3.12+, MIT, 328 stars.

  • kimi-k3-in-c

    FareedKhan-dev · Inference Engine

    Runs the 2.78-trillion-parameter Kimi K3 MoE on a single CPU in as little as 8.24GB of RAM, in portable C99 with no BLAS, no framework and no GPU. A streaming trunk architecture turns memory into a tunable dial and a packed expert cache evicts LRU. Output is byte-identical from 8GB to 224GB. Costs 1.7TB of storage and ~26.5s/token at 8GB. Apache-2.0, 3.9k stars.

  • claude-codex-settings

    fcakyon · Prompt Engineering

    Shared configs, hooks, agents and 50+ installable plugins spanning Claude Code, OpenAI Codex, Cursor and Gemini CLI. Covers code review, docs, database, deployment and payments, with auto-format hooks for Python, JavaScript, Markdown and Bash. Routes to Kimi K3, MiniMax and Z.ai GLM through Anthropic-compatible endpoints. Apache-2.0, 1k+ stars, 361 commits.

  • Plane

    makeplane · AI Infrastructure

    Open-source project management aimed squarely at Jira, Linear and ClickUp: work items with rich text, cycles with burn-down charts, modules for decomposition, filtered views, real-time analytics and AI-assisted Pages. Runs on React Router, Django, Postgres and Redis via Docker, Kubernetes or hosted cloud. AGPL-3.0, 55.7k stars, 5.3k forks.

  • Graft

    NanoNets · AI Infrastructure

    Builds a codebase into a knowledge graph of linked markdown nodes via tree-sitter parsing plus LLM summarisation, then feeds it to Claude Code, Cursor, Codex or Gemini so agents stop re-exploring. A 162-run sweep reports 46% fewer tool calls, 42% fewer tokens and 60% less latency; on 20 SWE-bench Verified instances, 75% correct against a 65% baseline. MIT, 1.5k stars.

  • authentik

    goauthentik · AI Infrastructure

    Self-hosted identity provider covering SAML, OAuth2/OIDC, LDAP and RADIUS, pitched as a replacement for Okta, Auth0, Entra ID and Ping. Deploys via Docker Compose, Kubernetes Helm, AWS CloudFormation or DigitalOcean. Now ships CLAUDE.md and an AI_POLICY.md alongside the code. MIT core plus a separate Enterprise licence, 24k+ stars, 23,586 commits.

Research & Safety (1)

  • WeatherNext Cyclones

    Google DeepMind · Research Paper

    Published in Nature on 6 August. Forecasts cyclone track, intensity and wind structure 15 days out using 1,000-member ensembles, adding more than 24 hours of lead time - three-day forecasts now match what two-day forecasts used to deliver. Runs a full 15-day forecast in under a minute on a TPU at 28x28km. Built with the National Hurricane Center and UK Met Office; weights open.

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