The archive · AI & Models · Technical decision · 2023
Alpaca.cpp bet a ChatGPT-style LLM can run on a MacBook; 10k stars
antimatter15's alpaca.cpp ran a ChatGPT-style LLM on a MacBook from 4 GB of weights; March 2023 Show HN, 673 points, 10.1k stars, upstreamed into llama.cpp.
alpaca.cpp (antimatter15)
What the business is
alpaca.cpp is an open-source program that runs a ChatGPT-style instruction-tuned LLM locally on macOS, Linux or Windows from a 4 GB quantized model, built by combining LLaMA, an open reproduction of Stanford Alpaca, and a chat-enabled llama.cpp.
How it started
On 2023-03-16, GitHub user antimatter15 posted alpaca.cpp to Hacker News as 'Show HN: Run an Instruction-Tuned Chat-Style LLM on a MacBook'. The project combined Facebook's LLaMA foundation model, an open reproduction of Stanford Alpaca's instruction tuning, and modifications to Georgi Gerganov's llama.cpp that added a chat interface; the headline demo ran on an M2 MacBook Air with 4 GB of weights.
What happened
The Show HN drew 673 points and 283 comments. The README positioned alpaca.cpp as a fast local alternative to ChatGPT-style APIs and shipped zips for macOS, Linux and Windows, while crediting the open chain behind it — LLaMA weights, alpaca-lora fine-tunes and llama.cpp's quantization. It also carried a research-only disclaimer: the weights were derivative of LLaMA, and Stanford Alpaca's instruction data had been generated by OpenAI, whose terms disallowed using its outputs to train competing models.
How it ended up
The standalone project was superseded: the README says the alpaca.cpp changes 'have since been upstreamed in llama.cpp' and directs users to llama.cpp instead. As of the 2026-09-05 fetch the repository remains public with 10.1k stars and 843 forks as a record of the demo.
Background
alpaca.cpp was a March 2023 experiment by GitHub user antimatter15 that ran a ChatGPT-style, instruction-tuned LLM entirely on a laptop: LLaMA foundation weights, an open reproduction of Stanford Alpaca's fine-tuning, and a chat interface built on llama.cpp. The demo was an unspeeded screencast on an M2 MacBook Air using 4 GB of weights, and the Show HN on 2023-03-16 drew 673 points and 283 comments.
The bet was that local inference mattered — that users would want a model that runs on their own machine with no GPU, no API key and no cloud, and that quantization plus a C++ runtime could make a 7B-class model usable there. The repo shipped zips for macOS, Linux and Windows with a one-command chat executable, crediting the open chain (LLaMA, alpaca-lora, llama.cpp) and warning that the weights were research-only derivatives of LLaMA and OpenAI-generated instruction data.
The standalone project did not become the long-term home of the idea: its changes were upstreamed into llama.cpp, and the README now says exactly that — 'Consider using LLaMA.cpp instead.' The repository remains public with 10.1k stars and 843 forks as of the 2026-09-05 fetch, a record of the moment local chat LLMs went from impossible to obvious.
What has to be true
- Timing: it landed days after Stanford Alpaca opened up instruction tuning, when demand for ChatGPT-like models was huge but everything still ran in the cloud.
- The technical shortcut was right: 4-bit quantization and llama.cpp's C++ runtime made a 7B chat model fit in 4 GB on a consumer laptop.
- Because it was a fork of llama.cpp, the improvements flowed back into the platform — which is why the demo's impact outlived the demo itself.
- The HN audience was exactly the people who wanted to run models locally, which is why 673 points and 283 comments came within hours of posting.
What can be applied
A sharp demo can prove a category before the platforms do: alpaca.cpp showed local chat LLMs were real in March 2023, but the durable win went to llama.cpp, which absorbed its changes.
Aftermath
As of 2026-09-05 the alpaca.cpp repository is still public under antimatter15 with 10.1k stars and 843 forks, and its README's first guidance is to use llama.cpp instead, noting that the changes were upstreamed. The project is effectively finished as a standalone fork, preserved as the artifact of the early local-LLM moment it helped demonstrate.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card