Qwen3.8-27B UD-IQ4_XS Heretic + MTP on a 16 GB card with 55-68 tok/s (24gb and 12gb versions available too)

I wanted the uncensored Qwen3.8-27B (llmfan46's Heretic build, MTP head preserved) on my RTX 4080 with MTP on (because I've been testing TONS of configs, and realized that would be the one)

And none of the published quants were built for that: the good IQ4_XS doesn't leave room for MTP, and the ones that fit are 3-bit.

I was amazed by that. as 16gb is VERY common, so decided to quantize it to fit, work well and loose as less as possible in quality:

I copied Unsloth's per-tensor UD recipe onto llmfan46's BF16, made 12 / 16 / 24 GB versions, and measured KLD against a Q8_0 of the same weights for every quant I could find.

Quality (mean KLD vs Q8_0, wikitext-2 / llama.cpp source code, lower is better)

Quant	GiB	Prose	Code
UD-Q5_K_XL (mine, 24 GB)	19.44	0.0045	0.0037
mradermacher i1-IQ4_XS	14.26	0.0202	0.0149
UD-IQ4_XS (mine, 16 GB)	13.27	0.0268	0.0192
llmfan46 Q3_K_M	13.48	0.0639	0.0456
mradermacher i1-IQ3_M	11.89	0.0649	0.0465
UD-IQ3_XXS (mine, 12 GB)	10.18	0.0904	0.0587
mradermacher i1-Q2_K	10.12	0.1551	0.1017

Speed on the 4080 (llama.cpp b11457, 15K-token prompt, 32K ctx, MTP 2 drafts, mean of 3 fixed seeds): UD-IQ4_XS does 55 tok/s on code and 50 on prose, against 29 / 29 without MTP. i1-IQ3_M is faster (70 / 54) at 2.4× the KLD, I prefer quality here over speed, but your call.

Things I didn't expect:

• 2 draft tokens beat 3. On the same seeds at 40K: 68 vs 56 tok/s on code, 55 vs 36 on prose. Fewer rejected drafts.
• 48K vs 40K produced the exact same tokens (identical draft-acceptance counts per seed) and was 22% slower. That's pure VRAM spill into shared memory, nothing else.
• The 16 GB edge is brutal. The same 40K config gave 68, 51 and 42 tok/s on code depending only on whether the desktop was holding 1.2, 1.4 or 1.9 GB of VRAM. Opening a Chrome window mid-generation took decode to 0.2 tok/s. If you're on Windows with a monitor on the same card, watch "shared GPU memory" in Task Manager.
• MTP head precision barely matters. q6_K head vs IQ3_S head: 83% vs 83% acceptance on code, 58% vs 54% on prose.

Repo with all three files, the commands, and the scripts (recipe extraction, dry-run verification, KLD, seeded speed bench): https://huggingface.co/codavidgarcia/Qwen3.8-27B-Uncensored-Heretic-MTP-UD-GGUF

The 12 GB and 24 GB context numbers are computed from llama.cpp's reported buffers, not measured on those cards. If you run them, let me know what you get!

Credit to llmfan46 for the weights, Unsloth for the recipes and mradermacher for the imatrix

Enjoy!

edit: added the KLD chart since a few people asked about the numbers, (lower is better)

https://preview.redd.it/pxp048081juh1.png?width=3200&format=png&auto=webp&s=5ee30548379fa9e7edce9e0242844656f2e89c07