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

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

Questions welcome.

Repo: ninfer-uncensored