Revibing nanochat's inference model in C++ with ggml
Recently I wanted to see if I could vibe some serious C++ code. The result is a C++ re-implementation of Andrej Karpathy's nanochat's inferencing part (https://github.com/karpathy/nanochat), built on top of ggml. Unlike llama.cpp, this isn't a standalone binary; it is a C++ library & Python wrapper designed to swap out some core classes within the nanochat pipeline. For playability, I’ve tried to keep the dependencies to a minimum: just ggml, nanobind, and gtest for unit tests. Features and limitations: - A drop-in replacement of nanochat’s `GPT` and `KVCache` classes. So…
What it does
In the maker’s words, at launch
Recently I wanted to see if I could vibe some serious C++ code. The result is a C++ re-implementation of Andrej Karpathy's nanochat's inferencing part (https://github.com/karpathy/nanochat), built on top of ggml. Unlike llama.cpp, this isn't a standalone binary; it is a C++ library & Python wrapper designed to swap out some core classes within the nanochat pipeline. For playability, I’ve tried to keep the dependencies to a minimum: just ggml, nanobind, and gtest for unit tests. Features and limitations: - A drop-in replacement of nanochat’s `GPT` and `KVCache` classes. So far I’ve only tested this with `chat_web.py`. You can see how it's integrated here: https://github.com/k-ye/nanochat/pull/1 - Supports CPU and GPU (Metal yes, CUDA probably?). - Handles PyTorch-to-GGUF conversion automatically on demand. - Only float32 is currently supported. Benchmark: On an M3 Max (Metal), throughput is roughly 1/3 that of the original PyTorch implementation. I haven’t profiled the code yet, but I suspect the bottleneck is the lack of bf16 support. Motivation - Writing meaningful (& fun) C++ again: I used to spend a lot of my day-to-day time in C++ while working at various tech companies. These days, opportunities to use it for personal projects are rare, as it’s often hard to find a use case where C++'s advantages truly matter. - Testing "Vibe Coding" capabilities: Most of my current work is in UE5. Ironically, Blueprints—which were designed to help non-coders—have become a bottleneck in the LLM era... Admittedly, the AI agent has generated some FOMO in me, and I wanted to see if AI could handle a lower-level C++ implementation of a complex system from scratch. - Understanding the LLM internals. Why nanochat? It hits the "Goldilocks" zone: popular enough to be relevant, concise enough to be educational, and practical enough to deserve a serious C++ implementation. If you’re like me — an infra guy from the old days who feels a bit threatened by LLM and/or AI coding — I think nanochat is a great reference. Tinkering with it however you like is a nice way to demystify the tech. I relied heavily on Claude Code (CC) for the implementation. Overall, I am both impressed and genuinely pleased with the experience. Happy to answer questions, hear feedback or further discuss AI coding!
Does the same job
all alternatives →
- TVTiny-vLLM – high performance LLM inference engine in C++ and CUDAMay 2026 · github.com · ▲205
- WMWe made our own inference engine for Apple Silicon2025 · github.com · ▲186
We wrote our inference engine on Rust, it is faster than llama cpp in all of the use cases. Your feedback is very welcomed. Written from scratch with idea that you can add support of any kernel and platform.
- IBIn-Browser Graph RAG with Kuzu-WASM and WebLLM2025 · blog.kuzudb.com · ▲158
We show the potential of modern, embedded graph databases in the browser by demonstrating a fully in-browser chatbot that can perform Graph RAG using Kuzu (the graph database we're building) and WebLLM, a popular in-browser inference engine for LLMs. The post retrieves from the graph via a Text-to-Cypher pipeline that translates a user question into a Cypher query, and the LLM uses the retrieved results to synthesize a response. As LLMs get better, and WebGPU and Wasm64 become more widely adopted, we expect to be able to do more and more in the browser in combination with LLMs, so a lot of…
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 27d ago · cactuscompute.com
Launched alongside, January 2026
the whole month →- IN
Hey HN! I wanted to share something I built over the last few weeks: isometric.nyc is a massive isometric pixel art map of NYC, built with nano banana and coding agents. I didn't write a single line of code. Of course no-code doesn't mean no-engineering. This project took a lot more manual labor than I'd hoped! I wrote a deep dive on the workflow and some thoughts about the future of AI coding and creativity: http://cannoneyed.com/projects/isometric-nyc
AI · Jan 2026 · cannoneyed.com
Automatic AI-powered code reviews the moment you open a PR
Dev tools · Jan 2026 · kilo.ai
