m4-max

Reproducible on-device LLM benchmarks for Apple Silicon (iPhone 17 Pro, M4 Max): Apple Core AI, MLX, llama.cpp, LiteRT-LM and Core ML on the same model and harness, every number with its quantization and capture session; hybrid Mamba-2 models (Nemotron-3 Nano, Granite-4.0-H, Falcon-H1) included.

conv: one command converts any file to any format on macOS. Routes to ffmpeg, sips, vips or pandoc automatically, with batch mode and image optimisation.

Cursor-Auto / Claude-tier-style serving for local GGUF models on Mac (M4 Max, 64 GB). FastAPI router fronts llama-swap + llama.cpp, classifying each request into a coder, planner, or uncensored-planner tier. OpenAI-compatible API, opencode integration, per-project subshell, one `llmstack` console-script.

Fast MLX port of ZeroEntropy zerank-2 cross-encoder reranker. 10x faster than PyTorch MPS on Apple Silicon. bf16, validated.

Recover the native MTP predictor missing from the 8-bit MLX Qwen3.8-27B-Uncensored package, build a BF16 sidecar, and reproduce a 15.59 → 48.75 tok/s controlled M4 Max result with MTPLX.