Uncensored versions of the Qwen3.5 language models have been released, with variants of 27 billion (27B) and 2 billion (2B) parameters.

Qwen3.5-27B

The 27B model is described as the main event, featuring 64 layers, a hybrid DeltaNet + softmax architecture, a 262,000 token context window, and multimodal capabilities. The author declares no refusals (0/465) and complete uncensoring. Several quantizations are available, including IQ2_M (8.8 GB), IQ3_M (12 GB), Q3_K_M (13 GB), IQ4_XS (14 GB), Q4_K_M (16 GB), Q5_K_M (19 GB), Q6_K (21 GB), Q8_0 (27 GB), and BF16 (51 GB). IQ quantizations use importance matrix calibration.

Qwen3.5-2B

The 2B version is presented as a proof of concept. While being a smaller model, it is stated that uncensoring did not degrade its quality. Again, no refusals occurred (0/465). Available quantizations are Q4_K_M (1.2 GB), Q6_K (1.5 GB), Q8_0 (1.9 GB) and BF16 (3.6 GB).

Both models include mmproj files for vision support. The author recommends using a recent build of llama.cpp and advises against using Ollama. A 35B-A3B version is planned for release in the future.