Qwen3.6-35B-A3B-oQ2.5e-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 Qwen/Qwen3.6-35B-A3B, a mixture-of-experts vision-language model with 35B total and 3B activated parameters and a hybrid layout of Gated DeltaNet and Gated Attention layers. It includes the language model, the vision encoder and the multi-token prediction (MTP) head.
The effective average is 3.114 bits/weight for the language backbone and 3.312 bits/weight for the complete checkpoint.
Quantization and Bit Distribution
oQ2.5e 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 2-bit, group 64 | 25.770B | 71.679% | 7.500 GiB | 2.50 |
| Affine 3-bit, group 64 | 6.442B | 17.920% | 2.625 GiB | 3.50 |
| Affine 4-bit, group 64 | 0.805B | 2.240% | 0.422 GiB | 4.50 |
| Affine 8-bit, group 128 | 0.129B | 0.359% | 0.124 GiB | 8.25 |
| Affine 8-bit, group 64 | 2.328B | 6.474% | 2.303 GiB | 8.50 |
| BF16 | 0.478B | 1.329% | 0.890 GiB | 16.00 |
| Total | 35.952B | 100% | 13.864 GiB | 3.312 |
Storage Breakdown
| Component | Logical weights | Storage |
|---|---|---|
| Language backbone | 34.661B | 13.490 GB / 12.563 GiB |
| Vision encoder | 0.447B | 0.893 GB / 0.832 GiB |
| MTP head | 0.845B | 0.503 GB / 0.468 GiB |
| Total | 35.952B | 14.886 GB / 13.864 GiB |
License
Apache-2.0, following the source model. See LICENSE.