Ollama 0.19 boosts Apple Silicon with MLX

This is a real infrastructure upgrade, not a cosmetic release: Ollama is making Apple Silicon feel like a first-class local inference platform again, and the cache work may matter almost as much as the raw benchmark gains for agentic workflows.

• –MLX plus Apple’s GPU Neural Accelerators should cut both time-to-first-token and steady-state generation latency on newer Macs.
• –NVFP4 support narrows the gap between local testing and production-style inference formats, which is useful for teams comparing outputs across environments.
• –Cache snapshots, reuse across conversations, and smarter eviction are exactly the kind of changes that improve Claude Code-style branching loops.
• –The preview is aimed at bigger machines with 32GB+ unified memory, so the win is strongest for high-end Apple Silicon users.
• –Focusing on Qwen3.5-35B-A3B coding workloads signals that Ollama is optimizing for serious local coding agents, not just casual chat.