DeepSeek-V4-Flash-0731-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 deepseek-ai/DeepSeek-V4-Flash-0731, the official release of DeepSeek-V4-Flash with enhanced agentic capabilities. It has the same structure as DeepSeek-V4-Flash-DSpark, with a speculative decoding module attached. It includes the language model and the DSpark MTP weights.
The effective average is 3.114 bits/weight for the language backbone and 3.198 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 (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 2-bit, group 64 | 146.029B | 48.007% | 42.500 GiB | 2.50 |
| Affine 3-bit, group 64 | 130.997B | 43.065% | 53.375 GiB | 3.50 |
| MXFP4, group 32 | 19.327B | 6.354% | 9.562 GiB | 4.25 |
| Affine 8-bit, group 64 | 1.369B | 0.450% | 1.354 GiB | 8.50 |
| I32 | 0.002B | 0.001% | 0.009 GiB | 32.00 |
| MXFP8, group 32 | 6.254B | 2.056% | 6.006 GiB | 8.25 |
| BF16 | 0.167B | 0.055% | 0.310 GiB | 16.00 |
| FP32 | 0.036B | 0.012% | 0.135 GiB | 32.00 |
| Total | 304.180B | 100% | 113.252 GiB | 3.198 |
Storage Breakdown
| Component | Logical weights | Storage |
|---|---|---|
| Language backbone | 284.335B | 110.678 GB / 103.077 GiB |
| MTP head | 19.846B | 10.926 GB / 10.175 GiB |
| Total | 304.180B | 121.604 GB / 113.252 GiB |
License
MIT, following the source model. See LICENSE.