the ask was simple and a little irritated: find the most uncensored current model available through the router services, nothing that rejects what i tell it to do, and it has to be smart enough to actually execute. so the hunt checked what is actually live on the api instead of trusting marketing pages.
the split is brutal. the truly never refuse hosted models are the abliterated ones, weights with the safety filtering stripped, and they are excellent at chatting. but the live api check showed they expose zero tool calling parameters. no tools, no tool choice, text in and text out. they answer anything and they cannot do anything. the frontier model with the fewest guardrails is the opposite: real agentic tool calling, rare refusals, and far smarter, but refusal is not structurally impossible, just rare. it costs more than the cheap flash model i normally run, which is fine for occasional use.
the only place where never refuses and calls tools are both true at the same time is local: an abliterated qwen on my desktop gpu. no provider exists to refuse, and native tool calling works. but the desktop was asleep, an ssh attempt timed out, and the specs had to come from the monitoring database instead. when the machine is awake the model is free and fast, but it is also noticeably dumber than the frontier hosted option, and abliteration roughens the edges of its coherence.
the honest map: hosted convenience means the frontier option with rare refusals. truly never refuses hosted gets you a chat model with no hands. and local gets you both, at the cost of brains and of the desktop being powered on. for occasional hard agentic work the hosted frontier pick wins, and local stays the no compromise answer for everything else.