Ollama 0.32.4 runs Laguna on Apple Silicon and speeds up Qwen3 MoE
The local-runtime update adds MLX support for Poolside's Laguna, sharpens speculative decoding, and fixes Qwen3 MoE for a 4–9% gain on M5 Max.
Copy markdown
Laguna now runs on Apple Silicon via MLX
v0.32.4 adds MLX-engine support for Poolside's Laguna (the 33B/3B open MoE) on Apple GPUs, so M-series Macs can serve it locally without a CUDA box. One more capable open model you can run fully offline on a laptop.
Qwen3 MoE decoding is fixed — and 4–9% faster on M5 Max
The release fixes Qwen3 MoE decoding when experts are quantized at different types, and speeds up packed gate/up projection for a reported 4–9% throughput gain on M5 Max. If you run Qwen3 MoE locally, updating is free performance.
Tighter speculative decoding for draft models
Ollama now quantizes a draft model's output head at the type you request when building speculative-decoding drafts, giving more control over the draft path that accelerates token generation. It lands on top of v0.32.3's fixes for stalled downloads and new B200 and Windows ARM64 GPU support.