DeepSeek-V4-Pro 9B Runs Uncensored Locally

The compelling story is compact local reasoning, not “fully unfiltered” magic—the refusal claims are self-reported, and removing guardrails does not guarantee stronger capabilities.

• –Heretic reports 3 refusals out of 100 prompts versus 100 for the base model, with a measured KL divergence of 0.0422.
• –DeepSeek-V4-Pro provides training outputs, not model weights; this is a distilled Qwen3.5 model rather than a hybrid.
• –The 5.63GB quantized file lowers the hardware barrier, but context length, RAM, and inference speed still determine whether it feels usable on a laptop.
• –Training focused on math and STEM, with no coding-specific data, so developers should not assume it matches dedicated coding models.
• –Unrestricted local models are useful for research and red-team testing but should be isolated and moderated before shared deployment.