An uncensored version of Qwen3.5-122B-A10B, named "Aggressive", has been released, aiming to provide unfiltered responses without personality changes to the model.
Key Features
- No Censorship: The "Aggressive" version is designed not to refuse any requests, offering a complete response without limitations.
- K_P Quantization: Introduces new K_P quantizations that, through model-specific analysis, preserve quality where it matters most, offering superior performance compared to standard quantizations with a limited size increase (5-15%).
- Multimodal Support: The model supports text, image, and video inputs.
- Extended Context: 262K token context window.
Technical Details
- The model has a total of 122 billion parameters, with approximately 10 billion active (MoE).
- Hybrid attention architecture: Gated DeltaNet + softmax (3:1 ratio).
- 48 layers.
- Several quantizations are available, including Q8_K_P, Q6_K_P, Q6_K, Q5_K_M, Q4_K_P, Q4_K_M, IQ4_XS, Q3_K_M, Q3_K_P, IQ3_M, IQ3_XXS, IQ2_M.
For those evaluating on-premise deployments, there are trade-offs to consider. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these aspects.
๐ฌ Comments (0)
๐ Log in or register to comment on articles.
No comments yet. Be the first to comment!