Qwen 3.5 4B IQ2_XS: +16.67% Reasoning Performance From Tensor-Level Allocation

I was finally able to replicate tensor level allocation outside the Gemma family. huggingface.co/ByteOtter/Qwen3.5-4B-CADA-IQ2_XS After the Gemma 4 12b, e4b and gemma 3 4b results, I attempted to expand into qwen and ran into a few walls. After 2 version updates and a slightly different approach, I was able to replicate the effect on Qwen. The result: BF16 reasoning: 78.125 Stock IQ2_XS + imatrix: 46.875 QLAB allocation + same imatrix: 54.688 That's +7.812 percentage points, or a +16.67% relative improvement in reasoning over the stock imatrix quant. Model sizes: Stock: 1,630,594,336 bytes QLAB: 1,637,318,816 bytes Difference: +0.412% Same idea as the previous runs. Build imatrix from category based corpus, measure damage, then redistribute precision at the tensor level while staying inside the same byte budget. This time, the optimal target landed at iq2_xs and the improvement was not isolated to reasoning. This is a category targeted quantization experiment, not a general model improvement. The focus is a successful transfer of the allocation mechanism. I've now observed successful allocation across dense, moe, qat, non qat, gemma and now qwen. There is no post training, lora, pruning or weight updates here. This is solely achieved by changing where precision gets spent inside the quant budget. Also, I've started posting progress here: x.com/byteotter I'm running Qwen 1.5 a2 7b next. After that i'd like to tackle larger models. Unfortunately at 1-3 attempts and hours of compute and storage, it gets expensive. If anyone wants to help fund compute for that run Im at buymeacoffee.com/byteotter My estimate for qwen 3.8 27b is $140 per run on a rented digital ocean gpu droplet and it may take a few attempts. The end goal is still the same: Give qlab a full precision gguf, select capabilities you want, have it auto level at the cliff and apply damage recovery tensor allocation to preserve as much of those capabilities as possible within the target byte budget. Charts are built from my data by chatgpt. TL;DR: On Qwen 3.5 4B IQ2_XS, tensor-level allocation improved held-out reasoning from 46.875 to 54.688, a +7.812 point / +16.67% relative gain over the stock imatrix quant at only +0.412% bytes. Eight of eleven evaluated suites improved, with regressions in knowledge QA, structured output, and coherence. Most importantly, this is QLAB's first cross-family replication outside Gemma.