Quanta AI Labs

Top AI News Weekly

Week 43 · Oct 12 – Oct 18, 2026

Open Weights Week. Meta released Muse Spark Open on 15 October, an 816B mixture-of-experts model and the first Muse model with downloadable weights, the same week StepFun opened the weights of Step 5 Preview, NVIDIA shipped Nemotron 3.5 Super VL, inclusionAI released the 560B Ling-3.1-flash under MIT and Liquid AI doubled the training data behind LFM2.5-24B-A2B. Image and audio followed the same pattern, with open weights for Black Forest Labs' FLUX 3 Image Klein, Fish Audio's Drama 3 and the Apache-licensed MOSS-Image. The closed side kept moving too: Gemini 4 Argon reached general availability on 13 October, Kling 4.0 launched with 30-second native clips, Krea 3 targeted fast consistent edits, and Meta announced Muse Wonder for long-running agent work. World models were the research theme of the week, with Matrix-Game 4.0 and NVIDIA's StreamRender-H3 both pairing a real game engine with a video model that repaints its frames, alongside multi-player and long-memory video generation. Coding agents kept picking up creative work, with new skills for explainer videos, motion graphics and style-matched animation.

50 launches and research drops that matter for enterprise AI builders—curated, tagged, and ready for your next roadmap sync.

New drops

50

Unique sources

49

Key themes

Video · Agents · Language

language models

Language Models

LLMs, reasoning models, MoE architectures, and language understanding systems.

Open Weight FrontierMeta

Muse Spark Open

Meta's first open-weight member of the Muse Spark family, released 15 October after the API-only Muse Spark releases up to 1.3 (covered in week 36). It is a natively multimodal mixture-of-experts model with 816B total and 42B active parameters and a 1M-token context, aimed at coding and agentic work. Weights are on Hugging Face (manually gated), SageMaker, Google Cloud Model Garden and Microsoft Foundry, with hosted access through OpenRouter, Together, Groq, Fireworks and others, and support in SGLang, vLLM, TensorRT-LLM, MLX, llama.cpp and Unsloth; Meta says it runs on two M5 Ultra Mac Studios. Released under the Muse Model License with limited commercial terms. Meta calls it the most capable US open-weight model but publishes no benchmark numbers yet.

View release ↗
Frontier Agent ModelMeta Superintelligence Labs

Muse Wonder

A new Meta model series for personal and professional agent tasks, announced 12 October in two sizes, Muse Wonder and Muse Wonder Full. It pairs Active Memory, which keeps user context across conversations, with Active Focus, which keeps work going on long, hard goals. Muse Wonder is available through the Meta Model API and Muse Code, Muse Wonder Full to Muse Code subscribers, and Muse Wonder will later power the Muse app, subscribers first. Meta claims state-of-the-art results on Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon but gives no scores in the post. No pricing or context length is stated. Follows Muse Spark 1.3 (week 36).

View release ↗
Frontier ModelGoogle DeepMind

Gemini 4 Argon

Since Week 41 2026: Gemini 4 Argon moved from a limited release for vetted cyber defenders to general availability on 13 October. It is now on the paid Gemini API tier through AI Studio and Android Studio, for enterprises through Gemini for business and Google Antigravity, and in the Gemini app and AI Mode for Google AI Pro and Ultra subscribers. Introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 after 30 November, with a 1M-token limit. Google reports 87% on ERQA, first place on the Vals Index and 0.7% attack success on Gray Swan prompt injection, all self-reported, and is widening Fairwind cyber access.

View release ↗
Open Weight ReleaseStepFun

Step 5 Preview

Since Week 40 2026: StepFun has released the weights of its agentic flagship, which was previously only on its API. The 600B-total, 27B-active sparse mixture-of-experts model with up to 1M tokens of context is now on Hugging Face and ModelScope, ungated, as FP8, BF16 and NVFP4 checkpoints, with the release landing around 15 October. It uses a new StepFun Model License 1.0 in the Apache style that requires Powered by StepFun AI attribution above 100M monthly users or US$20M revenue and a separate commercial licence for hardware makers and large model-as-a-service businesses. The current model card lists text and image input; StepFun's own benchmarks focus on coding, agentic work, knowledge work and finance.

View release ↗
Hybrid Linear MoEinclusionAI (Ant Group)

Ling-3.1-flash

An open model from Ant Group's inclusionAI, uploaded 13 October, with 560B total and 25B active parameters and a 1M-token context. The architecture is hybrid-linear: 63 KDA layers to 9 gated MLA layers, 512 routed experts with 8 active, and switchable thinking. inclusionAI points to results on Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE, GDPVal-AA v2.1 and MCP-Atlas, shown only as images on the model card, and to its use inside Claude Code, Kilo Code and OpenClaw. Day-zero SGLang support with MTP speculative decoding runs on 8 H200s or 4 GB300s, plus vLLM and OpenRouter. MIT licence.

View model ↗
Efficient Agentic MoELiquid AI

LFM2.5-24B-A2B

The successor to LFM2-24B-A2B (covered in week 9), released 13 October with the same 24B total and 2B active parameters and 40-layer convolution plus grouped-query attention hybrid. Pretraining doubles from 17T to 33T tokens, the vocabulary grows from 65K to 128K and context rises from 32K to 131K, and it now always reasons before answering after a five-stage post-training pipeline that includes agentic RL. Against Qwen3.6-35B-A3B, Liquid reports 48.09 versus 38.19 on LiveClaw and 96.75 versus 93.06 on IFEval but 51.0 versus 66.2 on SWE-Bench Verified, vendor-reported. It runs at 184 tokens per second on an M5 Max. API pricing is $0.10 input and $1.00 output per million tokens; LFM 1.0 licence.

View release ↗
Legal Domain ModelZ.ai

GLM-Legal

A legal-domain model from Z.ai, announced 12 October and built on GLM-5.3-Flash (covered in week 35), a 320B-parameter mixture-of-experts with 18B active parameters and a 1M-token context. Z.ai says it was mid-trained and then post-trained with reinforcement learning across tens of thousands of realistic legal environments, and reports 19.2% all-pass on Harvey's LAB benchmark across 24 practice areas, 12.5 points above the base model, plus parity with unnamed frontier models on LegalBench and PLawBench, all self-reported. There is no public API: access is through a consultation form, and it also appears in the AutoClaw desktop agent. No weights or licence announced. The blog renders only with JavaScript, so details come from news coverage.

View release ↗
vision & image

Vision & Image

Image generation, editing, restoration, OCR, and visual understanding.

Open Multimodal MoENVIDIA

Nemotron 3.5 Super VL

A vision-language version of the 120B-total, 12B-active hybrid Mamba-Transformer mixture-of-experts behind Nemotron 3 Super (covered in week 11), announced 15 October with base, post-trained and NVFP4 checkpoints on Hugging Face and ModelScope under the OpenMDW 1.1 licence. NVIDIA reports 80.8% on CharXiv chart reasoning (84% with a Python interpreter) against 63% for Nemotron 3 Nano Omni, 63.44% on Hermes Bench and 87.05% on LangChain Evals, above Nemotron 3 Ultra, and more than 600 output tokens per second on Baseten, all vendor-reported. It runs on vLLM, SGLang, Ollama, LM Studio and NIM.

View release ↗
Image EditingKrea

Krea 3

Krea's new image editing model, launched 15 October. Edits leave untouched areas alone and keep products, labels, logos and characters consistent across images; it takes up to six reference images per edit, accepts style references, and turns sketches and moodboards into finished photos, generating natively at 2K. Krea measured median time per edit, queue included, between 24 September and 1 October at under 8 seconds against 33 for Nano Banana Pro and 90 for GPT Image 2, self-reported. Pricing is $0.06 per 2K image and $0.03 per 1K, on krea.ai, the Krea API and partners including fal, Replicate, Together AI and OpenRouter. No open weights. Follows Krea 2 (weeks 20 and 26).

View release ↗
Open Image ModelBlack Forest Labs

FLUX 3 Image Klein

A 5.6B-parameter, open-weights member of the FLUX 3 Image family (covered in week 41) that runs on a single consumer GPU, with text-to-image generation, multi-target editing and multi-reference composition from one set of weights. It keeps the family's bounding-box layout control on a 0 to 1000 grid and ships as a distilled checkpoint for fast generation and an undistilled base for fine-tuning and research, with ComfyUI support. Weights are gated on Hugging Face under the FLUX Kommunity License, and product use needs a commercial licence. The repos went up on 13 and 14 October; BFL publishes no benchmarks or Klein-specific pricing.

View release ↗
Open Text-to-ImageOpenMOSS

MOSS-Image

An open text-to-image model family from the OpenMOSS team, announced 16 October and built on a 5B-parameter DiT backbone designed for efficient generation and fine-tuning, in Base and SFT checkpoints. Beyond ordinary prompts it supports layout-driven generation from structured JSON prompts, which can also drive controlled edits, and the team highlights text rendering inside images. It runs on a Diffusers-based pipeline with a flow-matching Euler sampler. Weights are Apache-2.0 on Hugging Face and ModelScope, with a technical report and user guide. No benchmark numbers on the page. Ostris's AI Toolkit added training support the same day.

View release ↗
Diffusion Fine-Tuningostris

AI Toolkit

Ostris's long-running toolkit for fine-tuning and training LoRAs on diffusion models, with a web UI and a self-updating launcher. This week it added FLUX.3 image training with multi-reference sampling and ComfyUI-format weights (15 October), MOSS-Image and MOSS Image Base support including edit-model and SFT edit training (16 October), and first support for fine-tuning and captioning with Qwen 3.8 27B (18 October), along with a patch for a path traversal issue in the captioner. There are no GitHub release tags; the version file reads 0.13.34. MIT, 12,349 stars.

View repo ↗
video & animation

Video & Animation

Video generation, editing, inpainting, and real-time rendering.

Text-to-VideoKling AI (Kuaishou)

Kling 4.0

Since Week 19 2026, when Kling Video 3.0 added a 4K mode: Kuaishou's Kling AI launched Kling 4.0 on 15 October, after a 4.0 Flash preview for top-tier annual subscribers from 28 September. According to the launch press release, the model generates up to 30 seconds natively in one pass, accepts prompts up to 8,000 tokens and up to 15 references (10 images, 5 videos, 7 subjects) with up to 10 keyframes, and works with Kling's MCP server and CLI so AI agents can drive it. The release lists 4K, 1080p, 21:9 ultrawide, 10-bit HDR, extensions up to two minutes and enterprise API access as coming later. Kling's own release notes render only with JavaScript, and coverage differs on the exact launch day.

View release ↗
Neural Game RenderingSkywork AI

Matrix-Game 4.0

Follows Matrix-Game 3.0 (covered in week 13) with a change of approach, released 15 October: an executable game engine now handles gameplay, physics and player input, and a causal few-step video model turns the engine's low-poly frames into detailed visuals chunk by chunk, which Skywork calls streaming neural rendering. It ships three playable scene models (Medieval Castle, Ancient Egypt, City Racing), a browser demo, game source code and single-GPU inference for A100, RTX 5090, 5080 and 5070 Ti, with the README reporting up to 67 FPS on an RTX 5080 and 58 on a 5070 Ti at 864x480, self-reported. Apache-2.0, technical report and weights on Hugging Face.

View project ↗
Streaming Game RenderingNVIDIA Research

StreamRender-H3

Part of NVIDIA's Sol-Engine work, after SoL-Refiner (week 41). A three.js racing game handles maps, physics and rules and streams a 13-class semantic map to a diffusion renderer built on MiniMax-H3 Ref2VA, which turns it into photorealistic video, with a reference image and prompt setting the look so one engine stream can be restyled. Training moves through teacher forcing, Resampling Forcing and distribution-matching distillation down to two denoising steps per chunk. Render time is 218 ms per streaming step on 8 GB200s, down from 799.7 ms, and a fast causal VAE decodes two latent frames at 720p in 23.8 ms. Code is in NVlabs/Sana; no end-to-end frame rate is stated.

View project ↗
Multi-Player World ModelHKU and Tencent ARC Lab

Multics

A generative world model, posted 12 October (arXiv 2610.17201), that renders one shared environment from several players' viewpoints. Each player supplies a starting view, a text prompt, actions and a camera path, and the generated videos stay consistent wherever views overlap. Rather than every player attending to every other, communication runs through player, group and global levels with shared group tokens and grouping based on recent camera paths. Trained on four players, it runs on 6, 8 or 10 without retraining, and at ten players the authors report 61.9% lower DiT latency and 86.1% fewer self-attention FLOPs than fully connected attention, self-reported. Apache-2.0 code and weights.

View project ↗
Long Video MemoryCornell, UC Berkeley, Harvard and Impossible Inc.

MosaiChunk

A memory method for autoregressive video generation past a frozen generator's context window. Instead of retrieving whole past chunks, it splits each chunk's key-value cache into sections with descriptors held in CPU memory; the newest chunk queries those descriptors and assembles the best-scoring sections into a mosaic within a fixed memory budget. Only the descriptor encoder is trained, by self-distillation with no ground-truth video. Tested on MiniMax-H3 and LingBot-World-Infinity with a new RememBench benchmark, it reports 0.899 CLIP with one chunk of memory against 0.841 for two-chunk whole-chunk retrieval, self-reported. Paper arXiv 2610.02153 (1 October), code 8 October.

View project ↗
Persistent World FilmmakingZhejiang University, Monash, UC Berkeley BAIR and Adelaide University

ATAS

Any Time, Any Space keeps one persistent world behind a multi-shot film. A compiler turns a screenplay into shot requests, each with a location, camera and event time, and before each shot it pulls context from a scene memory of static geometry in a shared frame and an event memory of time-indexed observations. Each generated clip is reconstructed and updates the world only if validation passes. Across four video backbones (MiniMax-H3, Wan 2.7, HappyHorse 1.0 and Wan 3.0) it posts the highest multi-shot average, for example 0.734 against 0.648 for ReCA on Wan 3.0, though it trails ReCA on VBench with Wan 2.7, self-reported. MIT code (repo created 12 October), arXiv 2610.20571.

View release ↗
Causal World StreamingTingting Liao

Streaming Code World Model

A streaming rebuild of Code World Model (week 35), released 12 October, that generates 0.71-second chunks causally from the start image, the proxy so far and earlier chunks with no access to future frames. It uses a sliding six-chunk attention window, diffusion forcing with a noisy history to slow drift, and 4-step DMD self forcing that runs 9x faster than 35-step sampling. On four held-out 17-second CARLA clips the 4-step model scores lower pixel error than the 35-step baseline. It takes 5.1 seconds per chunk on one H200 with about 108 to 123 GB of memory, and the author says it is not real time yet. LoRAs for MiniMax-H3 on Hugging Face; no licence file, results from four clips.

View repo ↗
AI Video EditorMartinDelophy

Timeline Studio

A browser-based, local-first video editor with a CapCut-style multi-track timeline that people and AI agents edit together, with models running in the browser on WebGPU or WASM: Kokoro and Piper voiceover, Stable Audio 3 Small music, Whisper captions, watermark removal, 4x restoration, vocal separation and a LivePortrait-based talking avatar. v1.0.9 on 13 October lets agents run relighting analysis and depth-video conversion through WebMCP tools with review and undo, and 17 October updates added in-browser denoising with RNNoise and DeepFilterNet3. The Anna cloud edition is a separate hosted product. MIT, 857 stars.

View repo ↗
Explainer Video AgentAlisa0808

Vox Director

An agent skill that turns a one-line topic into a finished Vox-style paper-collage explainer or ad: a beat map, a style bake-off, keyframe posters with Nano Banana 2, animation with Gemini Omni Flash or Kling O3 Pro, xAI narration, MiniMax music and ffmpeg assembly with captions, with two human approval gates before generation. It can also restyle a talking-head video or build collages around one still photo, and since 5 October adds a lip-synced on-camera presenter using Seedance 2.0. Runs on the paid Atlas Cloud API. A similar Vox generator under Anil-matcha appeared eight days after this one. MIT, 2,335 stars.

View repo ↗
Style-Matched Video Agentsedenfunf

ReelMimic

Give it a reference video, from a file, a phone recording or a YouTube link, and a one-line brief, and it makes a new 2D animated video in the same style. It first analyses shot count and lengths, tempo, transitions, colour, framing and camera moves, then shows a storyboard and style frames for approval, after which up to six Claude Code or Codex agents build shots in parallel, each reviewed by a different agent with before and after screenshots required for every fix. Seven drawing engines include HyperFrames motion graphics and watercolour. A 30 to 60 second video takes 1 to 3.5 hours on your plan. MIT, 3,633 stars, created 28 September.

View repo ↗
Animation EngineAcadema Labs

ManimGX

An animation engine for maths and 3D videos built for coding agents, released 7 October. It accepts Manim Community Edition's Python API but renders on a Rust and wgpu GPU core, installs with pip on Python 3.13 with fonts, typesetting and encoding bundled, can be set up by giving Claude Code, Codex or Cursor its llms.txt, and also runs in the browser. Academa says five 1080p60 scenes totalling 36.2 seconds rendered in 3.08 seconds on an M4 Pro, 90.6x faster than ManimCE and 11.5x faster than ManimGL, self-reported. v0.1.1, MIT, 1,821 stars.

View repo ↗
Math Animation3Blue1Brown

Manim

Moved from Week 8 2025: Grant Sanderson's Python animation engine for explanatory maths videos, used for the 3Blue1Brown channel and now a common render target for LLM-generated maths explainers. Scenes are written in code and render to video with precise vector shapes, LaTeX and camera moves. There has been no tagged release since v1.7.2 in December 2024; a September 2026 commit added VideoMobject and Sprite classes. This week brought a faster Manim-compatible engine, ManimGX. The community edition, ManimCE, is a separate project. MIT, 95,309 stars.

View repo ↗
Motion Graphics SkillTejashmakwana

motionmaxxing

An agent skill, released 6 October, for coding agents that make launch films, product promos, brand stings and kinetic-type videos, aimed at removing the slide-deck look of AI motion graphics. The author says 53 frame-accurate professional clips were broken down by six AI analyst agents, with human review, into about 150 designer rules, and the skill ships checks that fail work on corner labels, deck layouts, typed-out headlines and runs of dead frames. The README shows before and after versions of the same brief, which are the author's own comparisons. Early release, Apache-2.0, 875 stars.

View repo ↗
Seedance Prompt LibraryLearnPrompt

Awesome Seedance

A prompt library for ByteDance's Seedance 2.5 and 2.0 video models (covered in weeks 20 and 31) with 851 cases traced to their original posts, 495 cross-model retest runs, 27 copy-ready templates by clip type such as UGC creator reviews, and about 80 installable AI-video skills for coding agents. It is synced daily from the goodcase.ai case site, in English, Chinese and Japanese, and its templates are meant to be filled in, refined in any chat model, then pasted into Dreamina or Jimeng. Counts change daily. MIT, 2,139 stars, last synced 18 October.

View repo ↗
Facial Performance Maskssirioberati (TEIN AI)

Emotion Control

A pipeline, released 14 October, that pulls an actor's facial performance out of a video as a glossy clay mask, so a video model can redraw the face while keeping the expression. It tracks 478 landmarks and shot boundaries, fits expression and geometry with SPECTRE and FLAME, adds grayscale eye and mouth detail and crease shading, and renders a 512x512 mask MP4 that keeps the original audio; the examples feed these masks to Seedance with a separate appearance reference. No video model or paid API is bundled. The wrapper is MIT, but FLAME 2020 is research-only and SPECTRE is CC BY-NC-SA, so the full stack cannot be used commercially. 31 stars.

View repo ↗
audio & speech

Audio & Speech

TTS, STT, voice cloning, music generation, and audio processing.

Directable TTSFish Audio

Fish Audio Drama 3

Fish Audio's new flagship speech model, launched 13 October and following Fish Audio S2 (covered in week 11), built for directed performance. Delivery notes go in square brackets inside the script, seven paired tags (whisper, emphasis, soft, fast, slow, stress, forceful) change only the words they wrap, and multi-speaker scenes generate in one pass. Fish Audio reports 93.2% on Artificial Analysis' Pronunciation Robustness benchmark, first at launch, against 88.1% for its S2.1 Pro, self-reported. It runs on fish.audio, the mobile apps and the API, and open weights are on Hugging Face, free for commercial use below US$500,000 in annual revenue. Technical report arXiv 2610.18010.

View release ↗
Voice Cloning TTSKittenML (Stellon Labs)

KittenTTS 2

A 1.7B speech language model that writes codec tokens a vocoder turns into 24 kHz audio, a new model from the team behind the small KittenTTS (week 8). It clones a voice from 5 to 30 seconds of one speaker and ships 47 built-in voices, nine of them carrying Arabic, Chinese, French, German, Hindi, Italian, Portuguese, Russian and Spanish, with beta tags for ten emotions, ten vocal events such as laughs and sighs, and emphasis. The default weights are packed as 1.58-bit ternary (954 MB) and produce audio bit-identical to the 3.47 GB bf16 file, with GGUF files and a llama.cpp fork for CPU use. Stellon Labs Community License: commercial use is free only below $1M in annual revenue or funding. Published 30 September.

View model ↗
Local Dictation Appaltic-dev

FluidVoice 1.6.11

A free, open-source dictation app for macOS, at fluidvoice.com, that runs speech recognition on the device, with its Fluid-1 models cleaning up filler words and formatting and cloud AI providers optional; Windows and iPhone are in beta. Version 1.6.11 on 12 October adds FluidMeet, which records meetings locally with live captions, speaker labels and transcripts, plus app-wide search, Spoken Send, pronunciation training in the custom dictionary and smaller English-only Fluid Speech Mini and Pico models. The 1.6.12 beta on 13 October brings back meeting summaries, locally or through your own provider, and lets Claude Code and Codex read meetings. GPL-3.0, 12,283 stars.

View repo ↗
3d & spatial

3D & Spatial

3D generation, reconstruction, PBR materials, depth estimation, and spatial computing.

4D RiggingWestlake University, Nanjing University, HKU and MUST

GaussiAnimate

A SIGGRAPH Asia 2026 method, with its full code released 12 October, that rigs animatable subjects from reconstructed 4D Gaussian Splatting or mesh sequences, including clothed humans, quadrupeds, birds and garments. It combines a kinematic skeleton with free-form bones in a representation called skelebones, and creates new motion through Partwise Motion Matching, which retrieves and blends existing bone motions. The authors report 17.3% PSNR gains over Linear Blend Skinning on unseen poses, 21.7% over Bag-of-Bones and, with about 1,000 frames of data, 48.4% lower RMSE than robust skinning, self-reported. MIT code; paper arXiv 2604.08547 (April).

View release ↗
Image-to-3DPeking University, Math Magic and HKUST

UniForge3D

A model, posted 11 October (arXiv 2610.15357), that builds complete 3D assets from sparse, unposed images by reconstructing the geometry the inputs show and generating unseen parts from learned shape priors. Both branches work in one shared space that follows the anchor view's azimuth while keeping the object upright, removing the usual mismatch between camera-relative reconstruction and canonical-space generation. The branches exchange features to predict per-view point maps and a sparse voxel structure, and each voxel takes guidance from its most informative view. The authors report state-of-the-art appearance and geometry on Toys4K and GSO without giving numbers on the page. Code is listed as coming soon.

View project ↗
Feed-Forward Gaussian SplattingZhejiang University, Shenzhen University, Hello TransTech and Causa Robotics

InfiniSplat v2

A feed-forward 3D Gaussian Splatting reconstructor that builds scenes from any number of input views, with or without camera poses (arXiv 2610.20341). A dual-branch design freezes the geometry and learns only the Gaussian attributes, learned opacity suppresses artifacts from inconsistent points, and training uses scale-aligned ground-truth cameras with large-angle target views so renders hold up under big viewpoint changes. The project page shows qualitative comparisons against DA3-GS, WorldMirror 2.0, tttLRM and ReSplat and a feed-forward reconstruction of more than 1,000 KITTI frames. Code for v1 and v2 is on GitHub, updated 17 October, with a Hugging Face demo.

View project ↗
Video Motion RetargetingMaxLiu22

Motion2MixamoRig

An experimental pipeline that takes the motion of one person in a video, or the pose in a still photo, and applies it to an Adobe Mixamo-rigged 3D character, writing preview videos and a skinned .glb for Blender or Unity. It runs as a Python command-line tool or a Blender 4.2 extension with a one-click runtime install, needs a free SMPL-X model, and downloads about 5 GB of inference weights including GVHMR on first run. Its first releases, v0.1.6 and v0.1.7, came out on 17 October, the latter adding support for Mixamo rigs exported from Cinema 4D. Tested on Apple Silicon Macs only. Apache-2.0, 1,331 stars.

View repo ↗
agents & automation

Agents & Automation

Agent frameworks, orchestration, coding agents, desktop agents, and embodied AI.

Open Source HumanoidRoboParty

RoboParty RP1

An open-source humanoid robot platform published from mechanical CAD up to learned behaviour. Sub-repositories cover schematics, PCB layouts, Gerber files, a bill of materials and assembly guides, the ROMOMO joint modules and firmware, a dexterous hand over CAN-FD, a Linux BSP, a runtime SDK and robot descriptions in 24, 28, 30 and 42 degree-of-freedom variants. The training repo covers locomotion, motion imitation and behaviour foundation models alongside related ICLR and ICRA papers, with real-time control at 500 Hz. The overview repo is GPL-3.0 (sub-repositories carry their own licences), created 1 October and pushed through 15 October, 474 stars. No build cost given.

View repo ↗
Humanoid Table TennisShanghai Innovation Institute and The University of Hong Kong

PingpongDuo

A hierarchical controller that lets two humanoid robots play table tennis doubles as a team (arXiv 2610.20579). A team-level planner picks positioning targets and filters unsafe ones with learned risk estimates, separate footwork and striking skills handle movement and hitting, and a residual on the racket arm refines strikes after the robot moves. In simulation, ball-return success is 13.2% higher than the best prior baseline; on real hardware a 2.15 m clear-and-recover cycle takes 2.4 seconds at a peak of 2.23 m/s, the robots sustain more than ten alternating strikes, and average rallies run 4.87 against 3.79 with rule-based positioning, author-reported. No code or data mentioned.

View release ↗
Smartphone RoboticsOpenBot (ob-f)

OpenBot v0.9

An open-source platform that turns an Android or iOS phone into the brain of a robot car whose body costs about $50, with person following, point-goal navigation and trainable driving policies running on the phone, plus firmware, a Blockly playground and web and Flutter controllers. v0.9.0 on 7 October is the first release since March 2025, adding a Flutter controller settings drawer with tilt or on-screen driving and app languages including Hindi, and fixing ESP32 firmware, BLE setup, WebRTC streaming and navigation crashes; Bluetooth-gamepad and no-phone-mode firmware followed on 18 and 19 October. Started at Intel Labs, ICRA 2021 paper. MIT, 3,547 stars.

View repo ↗
Agent Web StandardW3C Web Machine Learning Community Group

WebMCP

The proposed web standard that lets a page register its own JavaScript functions as tools for AI agents through document.modelContext.registerTool(), each with a plain-language description and a schema, so agents act through a site's own code and interface instead of scraping it. Chrome opened an origin trial in Chrome 149, Edge has one in Edge 150, and ChatGPT Desktop supports it, with experimental support in Brave's Leo. Spec commits from 13 to 16 October settled concurrency, permission and re-registration rules, and TypeScript types ship as webmcp-types on npm. 4,671 stars. Unrelated to the earlier community WebMCP project archived in week 21 of 2025.

View repo ↗
Pentest Agent Systemvxcontrol

PentAGI 2.2

Since Week 8 2026: the autonomous multi-agent penetration-testing system has shipped v2.0, v2.1 and v2.2 (5 October). The latest adds MiniMax, Mistral and xAI for 13 provider types plus Azure OpenAI through the custom provider, failover to a second provider, and Anthropic access through workload identity federation. It fixes CVE-2026-14784, where the host Docker socket was mounted into sandboxes: workers now drop capabilities, the socket is not mounted by default, and an optional startup check tries 27 escape attempts. Seven search tools are merged into one web_search tool. MIT, with an Enterprise Edition sold separately, 31,167 stars.

View repo ↗
Autonomous Sales Agentethanplusai

Harvey

A self-hosted sales agent that runs in a loop: it finds local businesses through OpenStreetMap, reads their websites without model calls, scores them against your customer profile, writes three-email cold sequences that mention something specific about each business, sends them after you approve, and works the replies toward a meeting. Model calls go through your Claude Code subscription, or since v0.3.0 on 13 October your Codex subscription, instead of a pay-per-token key, and it throttles itself to leave quota for your own work, with a local dashboard for approvals. Check your provider's terms on automated subscription use. Unrelated to Harvey the legal AI company. MIT, 1,512 stars.

View repo ↗
App Cloning SkillsJakeschincariol

Replica Skill

Eleven free Claude skills, released 3 October, that rebuild an existing app's features clean-room style as a chain sharing one replica/ folder: recon of screens, flows and data model from public pages and store listings; architecture; design tokens; screen-by-screen build; backend with auth and payments; bug testing; a parity score against the original; and an entrepreneur step that turns complaints in reviews into fixes, then branding, landing page, store listing and deploy. It installs as a Claude Code plugin or plain skill folders and needs only Python 3.8. The author says it copies function, not code, logos or copy. From the author of the YouTube Agent Skill (week 41). MIT, 3,423 stars.

View repo ↗
Social Media SkillsJakeschincariol

Instagram and LinkedIn Agent Skills

Two free sets of Claude skills built on the same template as the author's YouTube Agent Skill (week 41). The Instagram set has 13 skills: Reels scripts from 26 scored hook formulas, research that ranks reels against each account's own median, a caption linter, carousels, stories, DMs and a 100-point profile score. The LinkedIn set (github.com/Jakeschincariol/linkedin-agent-skill) has 11: posts from 21 hook formulas, comments, replies, inbox triage and a weekly plan. Both include a humanizer that strips AI tells and scores the draft, and neither posts anything. MIT, 1,890 and 2,347 stars; released in September.

View repo ↗
Instagram Agent Skillssergebulaev

Instagram Skills

Nine Claude Code and Codex skills for Instagram marketing from the author of the LinkedIn Skills already in the archive, sharing the same voice engine: captions and Reels hooks in your voice, hashtag sizing, carousel planning, a weekly content plan, a niche scan that reads real data through Apify, and profile optimisation. Every draft has AI tells removed and waits for approval before it can be published through Publora, a third-party service. It installs as a plugin in Claude Code, Codex, Claude Desktop or claude.ai. Versions 1.0.30 to 1.0.32 shipped on 14 and 15 October. MIT, 503 stars.

View repo ↗
Isometric Explainer Skillwheresryan22

Anatomy

An agent skill that explains a technical idea by drawing it as a working physical object, so a cache key becomes a pin-tumbler lock, as interactive isometric SVG figures with optional WebGL shaders and 3D effects. It runs in Claude Code, Antigravity, OpenCode, Codex and MiniMax Code, installs with npx skills add and uses only Node's built-in modules. v1.0.0 launched 10 October and v1.1.0 on 12 October lets an agent publish a figure to a community gallery through a pull request, records time and tokens used and lists 38 fixes; a later commit added A4 and Letter PDF export. MIT, 16,218 stars.

View repo ↗
Web Design Skillnateherkai

scroll-craft

An agent skill, also a Claude Code plugin, for building scroll-driven websites with coding agents. It ships a scroll engine with video scrubbing, pinned sections and sideways rails, eight page layouts that rule each other out, a fingerprint check that makes each new build differ from earlier ones, and typography and spacing rules with a refuse list that bans gradient text, invented statistics and AI-purple gradients. A headless-browser pass checks for dead scroll, text that never fully appears and per-line contrast. v0.3.0 added a process drawn from ten approved client sites, and v0.3.1 on 30 September brought security and Windows fixes. MIT, 3,645 stars.

View repo ↗
Reverse Engineering Skillaicodeking

leanrea

A single 8 KB agent skill, released 17 October, meant to replace the REA reverse-engineering MCP server (covered in week 42) and its 207 tools. It tells a coding agent which plain command-line tool to use for each kind of target, from Electron asar, Mach-O, ELF and PE to Go build info, Android APKs, WASM, SQLite and pcap, with ready-to-paste recon commands and common traps. The author's benchmark on Claude Code with Sonnet 5.5, four tasks with three reps each, reports 38% less wall time, 35% lower cost and 60% fewer tokens than REA while scoring 74 of 75 against REA's 75, self-reported on one machine. MIT, 665 stars.

View repo ↗
developer tools

Developer Tools

SDKs, inference engines, data pipelines, monitoring, and AI infrastructure.

Diffusion Model TrainerNerogar

OneTrainer

Since Week 14 2025: this open-source GUI and command-line trainer for diffusion models has added support for Chroma (August 2025), Qwen Image (September 2025), Z-Image (December 2025), Flux.2 Dev and Klein (February 2026) and Ernie Image (May 2026), alongside FLUX.1, SD 1.5 to 3.5, SDXL, Stable Cascade, PixArt, Sana and Hunyuan Video. It supports full fine-tuning, LoRA and embedding training with masked training, EMA, aspect-ratio bucketing, multi-resolution training, auto-captioning and sampling during training. There are no tagged releases; the dates come from the merge commits. AGPL-3.0, 3,245 stars, last commit 10 October.

View repo ↗
Claude Code Accountsrealiti4

claude-swap 0.27

A Python command-line tool, cswap, that switches between several Claude Code accounts without logging out, rotates accounts automatically before a rate limit hits, shows usage in a live dashboard and runs accounts in parallel, for the Claude Code CLI and the VS Code extension. Version 0.27.0 on 14 October adds cswap import-usage, so one machine can poll the usage endpoint and share readings with others instead of each spending the same quota; 0.27.1 follows a Claude Code config-file change and 0.27.2 handles Console personal API keys and makes config writes atomic on Windows. A community tool, not an Anthropic product. MIT, 3,620 stars.

View repo ↗
research & safety

Research & Safety

Research papers, alignment, red-teaming, interpretability, and benchmarks.

Parallel Text GenerationStanford Hazy Research

Next-Concept Modeling

A non-autoregressive language modelling method from Christopher Ré and Azalia Mirhoseini's group, published 12 October, that generates text from coarse concepts down to fine detail. A multiscale tokenizer learns a hierarchy of discrete codes, a transformer predicts each level with all of that level's codes produced in parallel, and a decoder turns the hierarchy back into text. The authors report 10 to 31x faster generation than strong diffusion language model baselines with about 4 to 5x fewer training tokens, at competitive or better quality on LM1B, OpenWebText, WMT14 De-En, XSum and Sudoku, tested at research scale. Paper arXiv 2610.14162, code at HazyResearch/ncm.

View release ↗
Agent RL FrameworkQwen

QwenGyre

A framework from the Qwen team for online reinforcement learning on very long-horizon LLM agents, where one run can last hours, take hundreds of environment steps and use nearly 1M tokens. It moves GPUs between rollout and training without stopping live runs, and its trajectory processor rebuilds branching histories, scores partial progress and removes duplicate paths. The paper says that on Qwen 3.8 2.4T with 700K-token rollouts, NL2RepoBench rose from 52.5% to 58.5% in 48 steps, with speedups of up to 1.85x over colocated and 1.78x over asynchronous setups, all self-reported. Posted 27 September; follows the Qwen3.8 2.4T open weights covered in week 33.

Read paper ↗