# Running the uncensored Qwen3.8-27B (HauhauCS) on a 4090 at 262K context and ~130 tok/s — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/a5656c03e5a91fba

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-09 05:27 ingested 2026-10-10 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1x1d0u4/running_the_uncensored_qwen3827b_hauhaucs_on_a/)

## retained text



```
HauhauCS ships their uncensored Qwen3.8-27B as GGUF only. NInfer, a C++/CUDA wanted its own format.

Now the same model that ran at 91.6 tok/s / 131K under llama.cpp does:

* 262K context (the model's full native window)
* \~130 tok/s decode with MTP3, 70.8% acceptance
* 3,591 tok/s prefill on a 9K prompt
* Perplexity within 1.3% of the official artifact, so the conversion is clean
* Vision and tool calls still work

• E8 4Bit quant 

Converter + writeup here: [https://github.com/T-Crypt/ninfer-4090/pull/4](https://github.com/T-Crypt/ninfer-4090/pull/4)

Questions welcome.

Repo: [ninfer-uncensored](https://github.com/T-Crypt/ninfer-4090) 
```



document fetched 2026-10-10 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1x1d0u4/running_the_uncensored_qwen3827b_hauhaucs_on_a
