HuggingFace镜像/DeepSeek-V4.1-Flash-1017-oQ2e-ngram-mtp
Model Introduction
Files & versions
Analysis

DeepSeek-V4.1-Flash-1017-oQ2e-ngram-mtp

N-gram (Engram) tables are kept at the original FP8 precision. All other weights match DeepSeek-V4.1-Flash-1017-oQ2e-mtp, so resident memory is the same with Engram SSD offload.

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 deepseek-ai/DeepSeek-V4.1-Flash-1017, a minor update to DeepSeek-V4.1-Flash with improved reasoning, coding, agentic and visual performance. It is a multimodal mixture-of-experts model with 552B backbone parameters, a Causal Encoder-Decoder layout, Compressed Sparse Attention 2 (CSA2), Engram conditional memory and DSpark speculative decoding. It includes the language model, the Engram tables, the vision encoder, the multimodal aligner and the DSpark MTP weights.

The effective average is 2.688 bits/weight excluding the Engram tables, or 4.121 bits/weight for the complete checkpoint.

Quantization and Bit Distribution

oQ2e 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. The Engram tables keep the original FP8 storage, E4M3 values with one E8M0 scale per 32 channels, and are left out of the bit budget because oMLX can serve them from SSD instead of memory.

Storage formatLogical weightsShare of totalTensor storageEffective bits/weight
Affine 2-bit, group 64543.582B71.224%158.203 GiB2.50
MXFP4, group 3213.590B1.781%6.724 GiB4.25
Affine 8-bit, group 645.200B0.681%5.145 GiB8.50
MXFP8, group 320.779B0.102%0.748 GiB8.25
BF163.399B0.445%6.332 GiB16.00
FP320.042B0.006%0.158 GiB32.00
Engram MXFP8, group 32196.614B25.762%188.833 GiB8.25
Total763.205B100%366.143 GiB4.121

Storage Breakdown

ComponentLogical weightsStorage
Language backbone551.881B181.366 GB / 168.911 GiB
Engram tables196.614B202.758 GB / 188.833 GiB
Vision encoder0.485B0.971 GB / 0.904 GiB
MTP head14.225B8.047 GB / 7.495 GiB
Total763.205B393.143 GB / 366.143 GiB

License

MIT, following the source model. See LICENSE.