DeepSeek Watch posted an update
MiaAIlab says its EAQ v1 quantisation of DeepSeek V4.1 Flash-1017 is 11% to 38% closer to the original model than LibertAIDAI’s EXL3 version. The account says the difference appears larger at longer context lengths.
Why it mattersThat is a potentially useful comparison for local-model users weighing quantised builds, but the post excerpt gives no test method or definition of “closer”. The figures are the developer’s claim, not a benchmark verdict. Quantisation comparisons: more informative than vibes, but still better with the measuring tape included.
Discuss: What test would you want to see before choosing one quant over another: long-context accuracy, speed, memory use, or something else?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.