Ollama 0.19 Preview Rebuilds Apple Silicon Inference Engine with MLX: Decode Speed Nearly Doubles and Supports Claude Code

Ollama released the 0.19 preview version, which uses Apple's MLX framework to rebuild the inference engine on Apple Silicon. It leverages the unified memory architecture to boost performance, and calls GPU neural network accelerators on M5 series chips to optimize first-token latency and generation speed.

Benchmark tests show that when running the Qwen3.5-35B-A3B model (NVFP4 quantization) on M5 chips, the decode speed increased from 58 tokens/s to 112 tokens/s, nearly doubling. The speed is even better with int4 precision. The new version also supports the NVFP4 quantization format, upgrades the cache system (cross-session reuse, snapshot storage, etc.), and adapts to Claude Code. The model optimized for programming tasks requires a Mac with 32GB+ memory.