DeepSeek-V4.1-Flash-1017 Runs Uncensored on RTX 4090

The headline is less “a 4090 runs a giant model” than “careful quantization and offloading are turning workstation-scale MoE models into practical local tools.”

• –The setup still requires substantial RAM, fast NVMe storage, Linux, and specialized inference software, so it is not a plug-and-play consumer deployment.
• –Reported performance varies by quantization and prompt length; the model card shows roughly 22–26 tok/s decode and up to about 409 tok/s prefill in its documented tests.
• –The uncensored behavior comes from a community-modified derivative, so output policy, safety characteristics, and quality may differ from the official DeepSeek checkpoint.
• –MIT licensing and local execution make this attractive for developers who need privacy, experimentation, or offline coding workflows.
• –DeepSeek’s official model is itself open-weight and optimized for coding and agentic tasks, giving this quantization a strong base to work from.