gemma-4-26B-A4B-it-oQ8e-mtp
This model was quantized using oQ (oMLX v0.7.1) mixed-precision quantization with Improved oQe Quantization (jundot/omlx#4385).
A mixed-precision MLX checkpoint of google/gemma-4-26B-A4B-it, the mixture-of-experts model of the Gemma 4 family from Google DeepMind, with 25.2B total and 3.8B active parameters, 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-4-26B-A4B-it-assistant.
The effective average is 8.786 bits/weight for the complete checkpoint.
Quantization and Bit Distribution
oQ8e allocates bits by measured layer sensitivity, then rounds each group with an importance matrix (1024 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 8-bit, group 64 | 25.225B | 96.184% | 24.961 GiB | 8.50 |
| BF16 | 1.001B | 3.816% | 1.864 GiB | 16.00 |
| Total | 26.226B | 100% | 26.825 GiB | 8.786 |
Storage Breakdown
| Component | Logical weights | Storage |
|---|---|---|
| Language backbone | 25.233B | 26.821 GB / 24.979 GiB |
| Vision encoder | 0.573B | 1.143 GB / 1.064 GiB |
| MTP head | 0.420B | 0.839 GB / 0.782 GiB |
| Total | 26.226B | 28.803 GB / 26.825 GiB |
License
Apache-2.0, following the source model. See the Gemma 4 license.