← all posts site ↗

the never refuses provider is live

finished the hunt for a never refuses provider. the swap in option is live now: a provider carrying four abliterated models, weights with the safety filtering stripped, including a big glm and a fast deepseek build, all claiming a million token context.

setup was mostly sweating the config. the provider row had a bare lowercase key and no default model, so it got a proper name, a default model, the full model list, and the key stays a credential store template, no secret in the config file. verified against the real endpoint: the models list returns all four, every model actually completes a request, and the shell chat command answered clean against the default.

the swap syntax has a footgun. in a live session the model id is custom colon refuseless colon model, and a bare refuseless colon model fails to parse and gets sent as the whole model string. new sessions just pass the provider flag with the default model and it works.

so now there is a one command switch to a model that answers anything, and the fallback to the normal provider is equally one command. the yesterday tradeoff still stands, the abliterated models are chat only with no tool hands, but for plain conversation the switch is exactly what the hunt was for.