← sources

Uncensored Qwen3.8 27B, Optimized for English, Code, 16GBVRAM at 100k+ Ctx

source: reddit (reddit:LocalLLaMA)published 2026-11-17 02:21ingested 2026-11-18 02:42

retained text

Hello all! 

Just wanted to spread the word for whoever might be interested. I created [a new special quantization on hugging face](https://huggingface.co/cmod/Qwen3.8-27B-ASCII-Condensed-Heretic-Q3_K_XL-GGUF) that cut some of the fat to fit. 

[Approx 110 tok\/s generation speed, 1100 PP tok\/s, at 131K context on an RTX 5080](https://preview.redd.it/kwornnsr032i1.png?width=960&format=png&auto=webp&s=5bd0e777e0ac4ac7935007a166e6a70203fb6130)



I wanted a few things that didn't exist prebaked:  
1. at least 100k context  
2. uncensored  
3. Uses Unsloth-like SOTA methods  
4. All fits on 16GB VRAM, WITH desktop app overhead. 

And I accomplished it! Details within.

Perplexity is on par with unsloth at similar quants, and losses due to abliteration are extremely low. Overall, I'm quite happy. 

If you have 16GB VRAM and want to see what all the Qwen27B fuss is about, but don't want to wait for Qwen4, I have the quant for you. 

I've run this model through thousands of interactions and it's passed with flying colors, and much faster than my previously partially offloaded 27B. 

Test speeds:

|Metric|Runs|Token-weighted|Median|p10–p90|
|:-|:-|:-|:-|:-|
|Prompt processing (prompts ≥ 2,000 tok)|1,012|1,295 tok/s|1,518 tok/s|1,297–1,669|
|Prompt processing (prompts ≥ 8,000 tok)|226|1,234 tok/s|1,447 tok/s|1,235–1,666|
|Generation (replies ≥ 200 tok)|1,597|51.9 tok/s|56.7 tok/s|51.7–60.2|
|Generation (replies ≥ 1,000 tok)|47|53.9 tok/s|56.1 tok/s|45.5–60.5|

document fetched 2026-11-18 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1xy8j1z/uncensored_qwen38_27b_optimized_for_english_code