Open Source Radar — August 13: decks, video & a 14MB agent
The coding-agent parade paused and the daily board filled with makers instead: a deck generator that ships real PowerPoint files, a financial-market foundation model, Lightricks' one-model audio-video release, and a tool-calling agent small enough to live on a thermostat. Four projects worth your attention.
PPT Master (Python, ~46k stars) — An AI workflow that turns PDFs, documents, or topics into native PowerPoint decks: real slide masters, native shapes with working adjustment handles, transitions, animations, and data-backed charts and tables — not flat text boxes, not a filled-in template. It runs inside any agent-capable AI tool, keeps your files local, and can follow your own .pptx templates, with speaker notes that export as audio narration. The star count alone is the signal — roughly 46k in a couple of months — because it solves the exact complaint everyone has with AI slides: you couldn't edit the output. If your team lives in PowerPoint and your agents keep producing decks nobody can touch, this is the fix.
Kronos (Python, ~37k stars) — The first open-source foundation model trained on the "language" of financial markets: candlestick data from more than 45 global exchanges, tokenized into hierarchical discrete tokens and fed to an autoregressive transformer, with open checkpoints from 4M to 102M parameters. It was accepted to AAAI 2026, ships fine-tuning scripts (including an A-share example built on Qlib), and landed on today's trending board hard. You'd use it to forecast K-line sequences, adapt it to your own market data, or study how to build a domain foundation model from scratch — and the repo is explicit that its demo pipeline is research, not a production trading system.
LTX-2 (Python, ~8.8k stars) — Lightricks' open audio-video foundation model, billed as the first diffusion-transformer model that does synchronized audio and video in one architecture: one checkpoint, spoken dialogue, sound effects, and picture together, with a distilled 22B variant recommended for fast inference and a trainer for LoRA and full fine-tuning. The repo bundles pipelines for text-to-video, image-to-video, audio-to-video, keyframe interpolation, retakes, and a dubbing mode that rephrases while matching speaker identity and lip movement — plus ComfyUI integration. It's the strongest signal yet that the "video model with native audio" race now has an open contender you can run yourself, not just a hosted API.
Needle (Python, ~4.6k stars) — Needle 2 is a 45M-parameter tool-calling model that fits in a single 14MB binary and runs a full session in about 28MB of RAM, aimed at phones, wearables, smart-home hubs, and robots. It returns structured JSON calls with a calibrated confidence score per response — act above your threshold, escalate below it — and a built-in retrieval head picks the top five tools from a large catalog each turn, with a grammar that makes off-schema output impossible. The pitch is that it trades wins with models like FunctionGemma 270M while being 5x to 70x smaller. If you've wanted an agent on a device that never phones home, this is the first credible one that fits.
Worth watching this week: the frontier is splitting — big labs chase one-model-everything video while tiny specialized agents colonize the edge.
Would you rather run a deck-building agent, a 22B video model, or a 14MB agent on your phone? Tell us in the comments.
Sources: PPT Master (GitHub) · Kronos (GitHub) · LTX-2 (GitHub) · Needle (GitHub)
