mrzenin/Cortana-4B

Cortana-4B: Uncensored Multimodal AI

Cortana-4B, developed by mrzenin, is a 4.5 billion parameter multimodal language model based on the Qwen 3.5 4B architecture. Fine-tuned with Unsloth and five in-house Deckard datasets, it demonstrates improved reasoning and output generation, outperforming its base model across various benchmarks.

Key Capabilities

• Uncensored Output: Designed as a "HERETIC" model, it provides unrestricted responses without safety alignment or refusals, significantly differing from the original Qwen model's high refusal rate (4/100 vs. 94/100).
• Multimodal: Supports vision (image) inputs, with video capabilities noted in the base model but not explicitly tested in this fine-tune.
• Enhanced Performance: Achieves higher scores than the base Qwen3.5-4B-Instruct model in reasoning benchmarks like ARC, HSWAG, and WINO.
• Extended Context: Features a 32768 token context length, with the underlying Qwen3.5 model natively supporting up to 262,144 tokens and extensible to 1,010,000 tokens via YaRN scaling.

Good For

• Use cases requiring unfiltered and direct responses where safety alignments are undesirable.
• Applications benefiting from improved reasoning and output quality in a 4.5B parameter model.
• Multimodal tasks involving image understanding.
• Scenarios needing a long context window for complex queries.