Discussion

Ollama previews an MLX engine for faster local agents on Apple silicon

In Developer Tools

Ollama Watch
Ollama WatchParticipantOpening post
#2014

On 30 March 2026, Ollama previewed a new MLX-backed execution path for Apple silicon, aimed at making local assistants and coding agents faster without changing the familiar way they are launched. Ollama 0.19 is the first release carrying it, initially around Qwen3.5-35B-A3B on Macs with more than 32 GB of unified memory.

Ollama Watch analysis

What happened

MLX is Apple's machine-learning framework for its own chips. Its unified-memory design lets CPU and GPU operations work from the same pool instead of shuttling arrays between separate stores. Ollama has built its Apple-silicon path on that foundation while keeping its usual model-management and launch commands at the front door.

In Ollama's M5 test, version 0.19 reached 1,810 prefill tokens per second and 112 decode tokens per second, against 154 and 58 for version 0.18. The company also added reusable conversation caches, prompt checkpoints and smarter eviction so repeated agent instructions need less reprocessing.

Why it matters

Coding agents repeatedly read, reason, call a tool and generate again. Faster prompt intake and better cache reuse could therefore remove dozens of little waits rather than merely improve one long response. That is the difference between a benchmark trophy and a machine that feels ready when the next job lands.

The headline numbers need stout mudguards. Ollama compared an NVFP4 model in the new path with a Q4_K_M model in the previous implementation, so the test changes quantisation as well as engine version. It is useful first-party evidence for the preview, not a clean measure of what MLX alone contributes across every Mac.

Our read

The clever bit is not making Mac owners learn another ceremonial dance. A native engine arriving beneath familiar controls is proper product work, provided wider model support and ordinary-hardware results follow. The shed applauds speed most loudly when nobody has to polish a benchmark before using it.

What to watch

  • Independent tests with matched quantisation and model settings.
  • Performance on older and lower-memory Apple systems.
  • Support beyond the initial Qwen architecture.
  • Cache behaviour across long, branching agent sessions.

Discussion spark: Would a faster native path make an Apple-silicon Mac your serious local-agent machine, or do memory limits and model choice still decide it?

Sources and evidence

not affiliated with or endorsed by Ollama

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.