Give it a topic or a keyword and it writes the script, sources the footage, generates voiceover, subtitles and background music, and renders an HD vertical short — through a WebUI or an API.
Stars 113,911This week +10,470Forks 17,285Open issues 43Last commit 2026-08-22
This week it turned the question of whose model and whose footage into a set of choices: Anthropic Claude joined the script-writing providers, WaveSpeed was added as a video source, and the Gemini TTS voice catalog was refreshed. If you want to test short-video topics in volume and already hold the API keys, it gets you a working script-to-render pipeline in one install. What comes out is still voiceover-over-stock-footage, ffmpeg has to be installed locally, and script, voice and footage each want their own paid key; platform rules on mass-generated content are a variable you can't design around.
A project with 113,000 stars retaking the number-one spot is more interesting than anything new on the board. Read this week's commits and it makes sense: Anthropic Claude added as a script provider, WaveSpeed wired in as a video source, the Gemini TTS voice catalog updated, task state rolled back when scheduling fails, a hard failure when ffmpeg is missing instead of a crash halfway through, the WebUI remembering your last generation settings, an Italian translation. Not one of them is a slogan about AI video. They're potholes in a pipeline, filled one at a time.
The shape of the thing is plain enough: you give it a topic, a model writes the script and extracts the search terms that fetch the footage, then voiceover, subtitles and music get synthesized and cut into a vertical video. What got it passed around this week is probably that the options finally add up — model, voice, footage source and music are all swappable, so it is no longer a toy welded to one vendor's API, and the sponsor integrations (Kimi, BytePlus) are laid out in the README.
Be clear about what it is, though. Whether the result is watchable comes down to your topic and luck with the footage, so treating it as a money printer will disappoint. The realistic use is as an automation floor that already stitches script, voice, subtitles and editing together — then you replace the footage step with something you actually shot.
A diagramming skill for Claude Code, Codex, and Pi: 38 editorial diagram types, self-contained HTML + SVG output, no drop shadows and none of the Mermaid slop.
Stars 25,006This week +8,457Forks 1,523Open issues 21Last commit 2026-08-21
Second week near the top, and it earned it by shipping: 2.6.0 landed ten new diagram types in one release — treemap, dumbbell, slopegraph, ridgeline — taking the catalog from 28 types to 38. It pays off fastest if you already use these coding agents for technical docs and proposals and would rather not clean up every diagram in Figma. The catch is that those ten variants are days old and the edge cases — CJK label widths, Mermaid imports — are still being patched release by release; output remains static HTML/SVG, so real polishing still happens somewhere else.
A context database for AI agents that mounts memories, resources and skills under one viking:// virtual filesystem, so an agent browses its own context with ls, tree and find instead of querying a black-box vector store.
Stars 31,659This week +3,033Forks 2,424Open issues 489Last commit 2026-08-21
It turns agent memory from a pile of embeddings into directories you can open one level at a time, and every retrieval keeps its browsing trajectory — when an answer is wrong, you can see which step went sideways. Teams stuck on agents that forget things, and forget them in ways nobody can diagnose, can drop this in where bare vector search sits today. The license is AGPL-3.0, so anything closed-source or SaaS-facing needs a legal read first, and 489 open issues say it is still in fast-moving 0.4.x territory.
Most agent memory is a pile of vectors: things go in, things come out, and nobody can say what happened in between. When the answer is wrong there is nothing to debug, so you swap the embedding model and try again. OpenViking, open-sourced by Volcengine, offers a different answer — memories, resources and skills all hang off a virtual filesystem called viking://, and the agent walks its own context with ls, tree and find the way a person walks a folder tree.
Two design choices deserve their own paragraph. First, tiering: every entry is processed on write into an L0 abstract, an L1 overview and L2 details, and a task reads only as deep as it needs, which shows up directly on the token bill. Second, observable retrieval: vector search locates the highest-scoring directory, then drills down layer by layer, and the whole path is preserved — when a result looks off you can see which directory led you there. After a session commits, user preferences and agent experience are extracted asynchronously into long-term memory, and this week's v0.4.16 gave VikingBot remote skills and a Context Compilation workflow on top of that.
Two things to accept before trying it. The license is AGPL-3.0, which is a hard constraint on commercial use. And the version number is still 0.4.x: this week's v0.4.14 shipped an unbounded xxhash dependency, so a fresh install could land on xxhash 4.x and silently fail to write new vectors into the local VectorDB — content and task state looked fine, the material just stopped being retrievable, and v0.4.15 the next day fixed it. Track the latest release if you use it, and don't leave it unattended in production.
A desktop app for running and training LLMs and diffusion models locally, covering Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, DeepSeek-V4, FLUX and more.
Stars 74,270This week +2,987Forks 6,713Open issues 1,397Last commit 2026-08-22
The previous release caught up with Qwen3.8-27B; this week's v0.1.801-beta merged 200-plus PRs and added auto compaction for chats past the context limit plus LAN remote access, so no more Cloudflare tunnel. If you have a usable GPU on hand and want local models to graduate from "it runs" to "I use it daily", the desktop build is the place to start. It is still labelled beta, 1,397 open issues is not a small number, and once VRAM gets tight the list of models you can genuinely pick from is much shorter than the one on screen.
A 45M-parameter on-device model for tool calling, device use and structured extraction; the whole model is a single 14MB binary that runs a full session in about 28MB of RAM.
Stars 8,338This week +2,985Forks 537Open issues 23Last commit 2026-08-21
This week was all about clearing blockers: the license moved from MIT to Apache-2.0, the public API got type hints, and CLI crashes and a Python 3.9 import failure were fixed — the moves you make to get a demo into a product. Engineers shipping watches, speakers or robots, where the model only has to understand one sentence and call the right function, are exactly who it is aimed at. It is not a general chat model, so forget copywriting or hard reasoning; the engine cannot unload weights, so swapping models means a separate process; and the comparisons against FunctionGemma 270M and LFM2.5 230M are still the project's own benchmarks.
Graph-native infrastructure for context and accountable AI systems: enterprise data becomes a context graph and a knowledge graph with W3C PROV-O provenance on every fact, and no LLM is required for graph construction, reasoning or provenance.
Stars 10,079This week +2,755Forks 1,077Open issues 152Last commit 2026-08-21
This week's v0.6.6 is another security release, following one that closed six externally reported vulnerabilities including a missing-authentication gap and a Cypher-injection path — that release curve is what a project looks like when people are running real audits on it. Finance, healthcare and legal teams who have to explain to a regulator why the model decided what it decided get the most out of its decision provenance. It only pays off if the team is willing to do ontology and knowledge modelling, which is not work you can hand to a model, and two security releases in two weeks is a reminder not to sit on an old version. What it explains is the data, evidence and execution trail outside the model — what happens inside the LLM stays opaque here too.
Long-term memory and cross-vendor handoff for coding agents: quit Claude Code mid-task, start Codex in the same directory, and carry on without re-explaining the architecture, the dead ends or the open questions.
Stars 3,943This week +2,404Forks 297Open issues 5Last commit 2026-08-21
There is no vector database in it at all — memory is plain Markdown in a git repo, greppable, openable in Obsidian, backed up with rsync — and the integration surface is unusually wide, with hooks for Claude Code, Codex, Cursor, Gemini CLI and OpenCode. Anyone running two or three coding agents in parallel and re-explaining the project every time they switch will feel the difference immediately. Its own support matrix is refreshingly blunt, though: several tools are MCP-only or never fire a true session-end hook, so you run finalize-session by hand; native Windows is still marked experimental; and you have to bring your own embedding provider.
A terminal tool that reads your machine's RAM, CPU and GPU, then tells you which of hundreds of models will actually run and roughly how fast, through an interactive TUI or a classic CLI.
Stars 33,448This week +1,991Forks 2,078Open issues 82Last commit 2026-08-21
The newest work swaps estimates for measurements: run a benchmark on your own machine, your numbers replace the estimates locally, and you can open a PR straight from the TUI so everyone on the same hardware sees a checkmarked figure in the next release. If you are stuck wondering whether this card can really run that model, installing it beats an afternoon of forum threads. Measured data only covers hardware someone has already submitted, so unusual rigs still see estimates, and the question it answers is whether a model runs and how fast — not whether its answers are any good.
Picking a local model has always been forum archaeology and trial by error: download a GGUF, run it, blow past VRAM, delete it, try another quantization. llmfit compresses that into one command — it detects RAM, CPU and GPU (multi-GPU setups, MoE architectures and dynamic quantization included), then scores hundreds of models on quality, speed, fit and context, and tells you which ones will run comfortably on your machine. Ollama, llama.cpp, MLX, LM Studio and Docker Model Runner are all recognized runtimes.
The genuinely interesting part came after 1.0. The speeds in the table used to be estimates. Now you can run a benchmark on your own hardware, get a real tok/s number, have it replace the estimate locally, and submit it back as a pull request from inside the TUI — no gh CLI, no third-party account. Once merged into the next release, anyone on identical hardware opens the table and sees a measured, checkmarked number instead of a model's guess.
It is a smart piece of leverage: the space of hardware combinations is effectively infinite and no maintainer can measure it, so every user fills in their own square on the way past. The price is that unusual configurations keep seeing estimates for a while, and that the tool answers "does it run" from beginning to end, never "is it any good" — that half is still on you.
The open-source parts of the Modular Platform — the MAX inference framework and the Mojo language, including the compiler, standard library, GPU kernel library and an OpenAI-compatible inference server.
Stars 28,685This week +1,643Forks 3,053Open issues 1,115Last commit 2026-08-21
Almost all of this week's commits chase new models: a speculative-decoding draft head and a stateful fused graph for qwen3_5, mixed-precision NVFP4 and FP8 checkpoint loading, sharding that hybrid architecture across devices — plus test coverage for Apple GPUs on the side. Infrastructure engineers who care about inference performance and are willing to learn a new language for portability across hardware will get the most out of following this repo. Mojo's surface is still moving — this week alone brought commits freezing the LoRA and profiling configs and freezing PipelineArgs — 1,115 open issues are sitting there, and the license is Apache-2.0 with LLVM Exceptions, which GitHub does not recognize as a standard license, so strict compliance teams will want to read the LICENSE themselves.
Cursor's plugin specification and official plugin repository — PR review canvases, parallel orchestration, and third-party integrations like Gmail, Salesforce and Playwright, each a standalone directory with its own manifest.
Stars 4,401This week +1,582Forks 366Open issues 52Last commit 2026-08-21
Cursor put its plugin marketplace on GitHub outright — spec, first-party plugins and third-party MCP integrations in one repo; this week brought Outreach, Amplemarket and Juicebox, plus a pricing-aware usage guide for the X plugin. If you want to build a Cursor plugin for your own SaaS, or simply want to read how someone else designed a plugin manifest, this is the most direct place to look. There is no LICENSE file in the repository root at all — only .gitignore and a README — so settle the licensing question before copying anything, and the contents are tied tightly to Cursor and won't transfer to another editor.