gemma-4-31B-it-oQ6e-mtp

[!NOTE] Updated 2026-10-29: Re-quantized with the improved oQe quantization in oMLX 0.7.3. Please re-download if you have an earlier copy.

This model was quantized using oQ (oMLX v0.7.3) mixed-precision quantization with improved oQe (jundot/omlx#4930).

A mixed-precision MLX checkpoint of google/gemma-4-31B-it, the 30.7B dense model of the Gemma 4 family from Google DeepMind, with text and image input, configurable thinking modes and a 256K-token context. It includes the language model, the vision encoder and a multi-token prediction (MTP) head merged from google/gemma-4-31B-it-assistant.

The effective average is 6.700 bits/weight for the language backbone and 7.004 bits/weight for the complete checkpoint.

Quantization and Bit Distribution

oQ6e allocates bits by measured layer sensitivity, then rounds each group with an importance matrix (128 samples x 512 tokens) and a weighted least-squares refit of its scale and bias.

Storage format Logical weights Share of total Tensor storage Effective bits/weight
Affine 6-bit, group 64 27.641B 87.080% 20.916 GiB 6.50
Affine 8-bit, group 64 3.061B 9.643% 3.029 GiB 8.50
BF16 1.040B 3.278% 1.938 GiB 16.00
Total 31.743B 100% 25.883 GiB 7.004

Storage Breakdown

Component Logical weights Storage
Language backbone 30.697B 25.708 GB / 23.943 GiB
Vision encoder 0.576B 1.144 GB / 1.066 GiB
MTP head 0.470B 0.939 GB / 0.875 GiB
Total 31.743B 27.792 GB / 25.883 GiB

License

Apache-2.0, following the source model. See the Gemma 4 license.