9B model wins all 7 benchmarks against Qwen 27B, refusing only 6 out of 100 prompts
MIKE released an Apache 2.0 Qwen3.5 9B uncensored version: 256K context, native vision, 185 tokens/s on a single GPU. The page shows about 1.96 million downloads. The numbers are impressive, but you also need to look at the testing methodology.
A 9B model beat Qwen's own 27B on 7 benchmarks. The uncensored version refused only 6 out of 100 test prompts, while the original refused all 100. MIKE posted the model page link on X.
The model page name is very long: Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF. The author is DavidAU. Tags include Image-Text-to-Text, GGUF, English, and the license is Apache 2.0. MIKE listed the specs: 256K context, native vision, 185 tokens/s on a single GPU. The screenshot shows 1,300 likes and 1,965,614 downloads.
The intuitive metric for a so-called uncensored version is fewer refusals under the same set of prompts. MIKE's stock version means the unmodified original version. Here the two numbers are 100 out of 100 refused, while the modified version refused only 6.
These numbers are discussion-worthy for a direct reason. The parameter count is small, yet it beats the larger Qwen 27B on benchmarks; the refusal rate drops from 100/100 to 6/100. In the replies, some people were planning to test it that night, while others focused only on the 6/100.
The testing methodology also matters. The tweet did not elaborate on what the 7 benchmarks specifically are or how the prompt set was constructed. For 185 tokens/s, it did not give the GPU type, quantization level, or batch size. GGUF is a common local inference format; people who want to run it will first look at quantization, VRAM, and license. Treat this as a lead, not a purchase order.
MIKE's free model roundup article:
https://x.com/i/article/2090086161075113984
Original tweet:
https://x.com/mikenevermiss/status/2117109897976307976
发布时间: 2026-11-02 12:30