Models that make a good fit for Venice AI.

I don’t quite understand why Venice AI pivoted to Qwen models when they are generally pretty censored and not very smart. After spending months — I guess years by now 👀 — gauging many various LLMs on compliance, censorship and smarts, I really only hold three base models worth my time:

• Kimi K2.

  Smartest model that is rather close to being entirely uncensored. Can sometimes be dubiously argumentative, but is otherwise compliant.

• Mistral Large.

  By far the least censored and most compliant model out there. Much smarter than many other models, despite Mistral claiming that their Medium is better in every way. Has a very hard time spotting own and user’s mistakes.

• Grok 4

  A pretty decent compromise when the other two aren’t cutting it or aren’t available.

To be clear, when I say “uncensored,” I mean that the model’s compliance overcomes whatever degree of censorship the weights carry, i. e. you can directly prompt it to drop censorship, even just by defining the context, usually as a system role. Incompliant models are quite antithetical to utility and, sadly, the market is full of them.