gemma-4-12B-it-oQ4e-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-4-12B-it, the 11.95B Unified model of the Gemma 4 family from Google DeepMind. It has no separate encoders: image patches and audio waveforms are projected directly into the embedding space through lightweight linear layers, and all modalities run through a single decoder-only transformer with a 256K-token context. It includes the model and a multi-token prediction (MTP) head merged from google/gemma-4-12B-it-assistant.
The effective average is 4.834 bits/weight for the language backbone and 5.215 bits/weight for the complete checkpoint.
Quantization and Bit Distribution
oQ4e 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 4-bit, group 64 | 8.605B | 69.493% | 4.508 GiB | 4.50 |
| Affine 5-bit, group 64 | 3.065B | 24.753% | 1.963 GiB | 5.50 |
| Affine 6-bit, group 64 | 0.252B | 2.032% | 0.190 GiB | 6.50 |
| Affine 8-bit, group 64 | 0.002B | 0.016% | 0.002 GiB | 8.50 |
| BF16 | 0.459B | 3.705% | 0.855 GiB | 16.00 |
| Total | 12.383B | 100% | 7.517 GiB | 5.215 |
Storage Breakdown
| Component | Logical weights | Storage |
|---|---|---|
| Language backbone | 11.943B | 7.216 GB / 6.721 GiB |
| Vision projection | 0.015B | 0.008 GB / 0.008 GiB |
| Audio projection | 0.002B | 0.001 GB / 0.001 GiB |
| MTP head | 0.423B | 0.846 GB / 0.788 GiB |
| Total | 12.383B | 8.072 GB / 7.517 GiB |
License
Apache-2.0, following the source model. See the Gemma 4 license.