Mike Hutu

2/2 No CUDA, no massive model downloads—just pull the repo, fire the Docker image, and you’ve got a chat‑ready bot on your edge node. github.com/louisho5/picob…

🧵 1/2 If you’ve ever cursed at the endless Docker‑compose files and GPU‑driver gymnastics required to get a bot up and running, meet Picobot. It’s a single 9 MB Go binary that spins up a fully‑functional AI agent in seconds.

3/3 github.com/MoonshotAI/Kim…

2/3 That's the kind of scale that used to be the exclusive domain of big labs, and it means you can finally get high‑fidelity speech‑to‑text, voice generation, and even multi‑turn audio conversation from a single open‑source checkpoint. 🎧

🧵 1/3 MoonshotAI just dropped Kimi‑Audio, a 7B‑parameter model trained on a staggering 13 M hours of mixed audio‑text data.

2/2 The new tmux‑Orchestrator repo wraps Claude agents in a resilient loop, auto‑restarting them whenever they exit and keeping the whole thing alive across SSH sessions. No more manual "screen -r" gymnastics. github.com/Jedward23/Tmux

🧵 1/2 If you’ve ever watched a Claude‑powered bot die in a tmux pane and then spent the next ten minutes hunting down the crash, you’ll love this.

2/2 For anyone tired of brittle OCR pipelines, this is a game‑changer – the model ingests the image and produces structured text in one go. github.com/X-PLUG/mPLUG-D…

🧵 1/2 The new mPLUG‑DocOwl repo is turning heads: a modular multimodal LLM that reads PDFs, scans, and charts without any OCR pre‑step. The codebase is pure Python, Apache‑2.0, and already has 2,378 stars on GitHub.

2/2 No more cobbling together fragile pipelines or paying for custom training—just pull the repo, load the model, and you’re ready for Mandarin, regional dialects, English, and even karaoke‑style input. github.com/FireRedTeam/Fi…

🧵 1/2 The FireRedTeam just dropped FireRedASR, an Apache‑2.0 Python stack that hits SOTA on public Mandarin benchmarks and, weirdly enough, can transcribe singing lyrics.

2/2 Drop in your model, pick your target (GPU, VPU, CPU) and the code swaps the right runtime under the hood. No more per‑backend forks, no more mysterious crashes. github.com/ankandrew/fast

🧵 1/2 Tired of juggling TensorRT, OpenVINO, and custom ONNX tweaks just to read a license plate? A new open‑source repo bundles a fully‑customizable ONNX pipeline that abstracts away the backend pain.

2/2 The new ai‑performance‑engineering repo from cfregly does the heavy lifting automatically – it profiles, rewrites, and tunes kernels on the fly, delivering reproducible speedups across both training and inference. github.com/cfregly/ai-per…

🧵 1/2 If you’re still hand‑crafting CUDA kernels to squeeze out a few percent on PyTorch, you’ve just wasted a lot of time.

2/2 No more manual prompt engineering—just type a comment, watch the code appear. github.com/breschio/drawb…

🧵 1/2 Ever wish you could skip the copy‑paste dance between design notes and LLM prompts? Drawbridge does exactly that. It watches your browser UI comments, turns them into Claude Code prompts on the fly, and feeds them straight into your coding agent.

2/2 No more Python loops, no more waiting for hours – you get results in seconds, and you can iterate fast enough to stay ahead of the curve. github.com/polakowo/vecto…

🧵 1/2 If you’re still looping over each strategy one‑by‑one, you’re losing money before the market even opens. The new vectorbt engine lets you backtest thousands of ideas in a single pass, thanks to pure‑NumPy vectorization.

2/2 The new getAsterisk/claudia repo flips the script: a dedicated GUI plus a pluggable agent framework that lets you spin up Claude sessions, swap providers, and keep your workflow in one place. github.com/getAsterisk/cl…

🧵 1/2 If you’re still juggling copy‑paste, fragile CLI flags, and manual session juggling to run Claude Code, you’re not alone.