Ollama v0.32.4: Apple GPU support via MLX and 4-9% speed boost
Local AI runner now runs on Apple Silicon GPUs with up to 9% faster inference.
Deep Dive
Ollama released v0.32.4, adding support for Laguna on Apple GPUs via the MLX engine. The update quantizes draft-model output heads at the requested type when creating speculative-decoding drafts. It also fixes Qwen3 MoE decoding for differently-quantized experts and includes faster packed gate/up projection (~4–9% on M5 Max). Available now on GitHub.
Key Points
- New MLX engine support lets Ollama run on Apple GPUs (Apple Silicon) without CUDA.
- Draft-model output heads are now quantized for more efficient speculative decoding.
- Qwen3 MoE fixed for differently-quantized experts; packed gate/up projection 4-9% faster on M5 Max.
Why It Matters
Makes local LLM deployment viable on Macs, boosting performance for Apple-based AI workflows.