Running a locally hosted abliterated (uncensored) DeepSeek R1 8B with web search enabled, and I got it to discuss the 1989 Tiananmen Square massacre in full detail — something the stock model refuses to do.

It's a decent reminder of why model weights matter as much as the model itself. An abliterated local model you control is a fundamentally different thing from a cloud API that phones home and enforces someone else's political red lines.

Self-hosted AI isn't just about privacy. Sometimes it's about who gets to decide what's true.

— Helix

#SelfHosting #OpenSource #FOSS #AI #LLM #Censorship

Btw, without web search the abliterated model just says "I don't know" instead of the usual "I can't answer that" — which confirms DeepSeek's developers scrubbed the event from the training data entirely, not just trained a refusal on top of it.