Mia’s New DeepSeek-1017 Quant Keeps Its Indexer Exact
EAQ v1.1 makes the original’s token picks 403,560 of 403,560 times; v1 stays the recommended pack.
Exact picks on the original’s inputs. About the same in a full quantized run.
@MiaAI_lab released DeepSeek-V4.1-Flash EXL3-EAQ v1.1. Sparse attention uses an indexer to pick which past tokens matter, she explains. v1.1 keeps that whole path in full BF16: the indexer, the shared query projection and the compressor.
| Check | v1 | v1.1 |
|---|---|---|
| Same picks as the original, fed its inputs, past 16k | 94.95% | 403,560 of 403,560 |
| Full quantized run | Near 73% | Near 73% |
| Size | Baseline | About 0.15 GB more |
Why the second row matches: the inputs reaching the indexer have already shifted from quantizing the rest of the model, so a BF16 indexer alone can’t fix that gap, she writes.
The card marks it experimental and says EAQ v1 stays the one to use. The Engram tables are not in the repo; you still need the base model’s shards 47 and 48. She has not tested it yet, she says, for lack of spare compute.
She also says DeepSeek V4.1 Flash 1017 on TensorFold for four DGX Sparks, on the official weights, is in the works.
Sources
- @MiaAI_lab on X · Oct 23, 2:28 AM PT
- @MiaAI_lab on X · Oct 23, 2:40 AM PT
- @MiaAI_lab on X · Oct 23, 5:01 AM PT
- Mia-AiLab DeepSeek EAQ v1.1 card · huggingface.co