DavidAU releases MN-26B-Oblivion, a 26B uncensored Mistral fine-tune for reasoning tasks

DavidAU released MN-26B-Oblivion-Uncensored on HuggingFace this week—a 26-billion-parameter text-generation model that removes safety tuning from a Mistral base. The model card tags it as "uncensored" and "not-for-all-audiences," with a focus on "instruct-reasoning," signaling unrestricted output and emphasis on multi-step reasoning. DavidAU built the fine-tune using Unsloth, a memory-efficient training framework that accelerates LoRA and full-parameter fine-tuning on consumer GPUs.

The weights ship as safetensors, the standard format for fast loading and reduced memory overhead during inference. At 26 billion parameters, the model sits between the 13B and 70B tiers common in the Mistral family, targeting users who need stronger reasoning than smaller models but cannot fit 70B checkpoints on a single card. The "instruct-reasoning" tag suggests the fine-tune emphasizes chain-of-thought and structured problem-solving over raw completion speed.

DavidAU has released other uncensored Mistral derivatives in recent months, building a catalog of abliterated checkpoints for practitioners running local inference without content filters. The Oblivion variant name implies a focus on removing residual safety layers rather than adding domain-specific knowledge. No benchmark numbers or example prompts appear on the model card yet, so real-world performance against other 26B uncensored checkpoints remains unclear. Early user evaluations and quantized GGUF releases will likely surface in the coming days—watch for community feedback on whether the reasoning gains justify the parameter bump over smaller abliterated alternatives.