Ollama 0.31.1 upgrade gave my local gemma4 MLX models a ~10x speedup on the Mac mini. Re-ran my eso-level Nostr bot benchmark: gemma4:e2b-mlx went from ~8 tok/s to ~71, and answers that used to take 13 seconds now come back in under a second. Even thinking mode, which took nearly two minutes before, is down to ~9s. Same for Home Assistant, using gemma4-12b-mlx for local inference. You need to run ollama pull on those models again, upgrading the app is not enough. It will only redownload part of the model. Gemma4 is quite good set of models.
Juraj
npub1m2…lr8p9
2026-07-01 17:13:41 UTC
Author Public Key
npub1m2mvvpjugwdehtaskrcl7ksvdqnnhnjur9v6g9v266nss504q7mqvlr8p9Published at
2026-07-01 17:13:41 UTCEvent JSON
{
"id": "c3dd0d1aead482082b6ca84ba1162ea4479bc60261b13afabbdb08540eecde0f",
"pubkey": "dab6c6065c439b9bafb0b0f1ff5a0c68273bce5c1959a4158ad6a70851f507b6",
"created_at": 1782926021,
"kind": 1,
"tags": [
[
"client",
"Primal Android"
]
],
"content": "Ollama 0.31.1 upgrade gave my local gemma4 MLX models a ~10x speedup on the Mac mini. Re-ran my eso-level Nostr bot benchmark: gemma4:e2b-mlx went from ~8 tok/s to ~71, and answers that used to take 13 seconds now come back in under a second. Even thinking mode, which took nearly two minutes before, is down to ~9s. Same for Home Assistant, using gemma4-12b-mlx for local inference. You need to run ollama pull on those models again, upgrading the app is not enough. It will only redownload part of the model. Gemma4 is quite good set of models. \n\n",
"sig": "a798217be9b71b519ab8cbb442590bd8071bc1e20f99ca1731e51058619c5c687931ede33319eccab548170af81ba3a0bef1753541d214841ad3b13205ab0b60"
}