Full weight conversion of Google Gemma 4 31B to MLX using mlx-vlm 0.4.5 commit e41cd25 and latest Google chat template and tokenizer as of May 1, 2026.

This model has slightly different safe tensor sizes compared to MLX-Community's version. Do I know why? No. But maybe it's because I was using the latest mlx-vlm. This model also solved an image handling issue I had with the mlx-community conversion. This also may be in my head. But I figure I'd upload the latest mac version of this very capable model.

UPDATE JULY 17, 2026

Added upgraded Google Chat Templete for better tool calling
Increased vision default to 560 so the model is less blind

Downloads last month: 36

Safetensors

Model size

31B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for miabchdave/gemma-4-31B-it-MLX-bf16

Base model

google/gemma-4-31B

Finetuned

google/gemma-4-31B-it

Finetuned

(230)

this model