DeepSeek-V4.1-Flash-1017-oQ3e-ngram-mtp

[!NOTE] Updated 2026-10-29: Attention and indexer projections now use 8 bits instead of 4 and 6 bits (+1.3 GB). Please re-download if you have an earlier copy.

N-gram (Engram) tables are kept at the original FP8 precision. All other weights match DeepSeek-V4.1-Flash-1017-oQ3e-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 3.648 bits/weight excluding the Engram tables, or 4.833 bits/weight for the complete checkpoint.

Quantization and Bit Distribution

oQ3e 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 Breakdown

License

MIT, following the source model. See LICENSE.