airllm

Run OpenClaw on a local model. Zero API cost. One OpenAI-compatible endpoint over Ollama, llama.cpp, vLLM or LM Studio — plus AirLLM for models bigger than your GPU.

Run 70B+ LLMs on a single 4GB GPU — no quantization required.

OpenAI-compatible API server wrapping AirLLM (github.com/lyogavin/airllm) — run large LLMs on small GPUs, CPU, Apple silicon or Docker

Custom ML architectures, training pipelines, and Dockerized deployment experiments — built from scratch. Ongoing.

This bridge was made to help developers use @AirLLM for ClaudeCode, Cursor and other agentic tools.

Локальный AI-агент для iPhone с layer-wise offloading и Apple MLX

Windows desktop shell for Original Kimi, standalone GGUF runtimes, and the AirLLM Kimi K3 hardware proof.

PyTorch inference acceleration engine: rejection-sampling speculative decoding (1.92x speedup), 5-way quantization bake-off & AirLLM 70B streaming on RTX 4060.