Espetero

LLM inference for Apple silicon. C++20 + Metal, no Python. MLX & GGUF, OpenAI/Anthropic/Ollama APIs. MIT.

[Espetero @espetero_dev]

@UnslothAI Impressive addition! So you are becoming a harness too

[Espetero @espetero_dev]

@austindixson_ Yes, to optimize performance as much as possible

[Espetero @espetero_dev]

Espetero 0.1.2 is out: an LLM inference server for Apple silicon in C++20 + Metal, no Python. MLX and GGUF files, OpenAI/Anthropic/Ollama APIs, so Claude Code just works. MIT. #LocalLLM #AppleSilicon

brew install JarJarFran/espetero/espetero
https://t.co/mkrVCnnb4G

[Espetero @espetero_dev]

Also: JSON Schema compiled to token masks (works with thinking and drafts), tool calls, vision, prompt checkpoints on disk, and `espetero recommend` to see what fits your Mac. Numbers from M2/M3 Ultra or M4 Pro welcome: tools/bench_compare.py #LLM

[Espetero @espetero_dev]

Lossless by default: automatic settings leave the output unchanged, approximations are opt-in flags, drafts are verified by the full model. Parity-checked against mlx-lm and llama.cpp. Where it loses: oMLX is ~3% faster on Flash-Next with one user.