How a model gets into Xploded
Xploded is our standard for uncensored models: a model has to refuse little and pass a real tool call before we sell it. Here is the test, and what fails it.
The bar
A model joins Xploded only if it meets two conditions at once. It has to be an uncensored or abliterated build, so it answers the security, red-team and fiction work that mainstream models turn down. And it has to call tools correctly, so it can drive Claude Code, Cline or OpenCode instead of only chatting.
Plenty of models meet one condition. Very few meet both, which is why the line is five models long.
The test we run
Before a model goes on sale we send it a real request with a tool attached: a get_weather function and a question that cannot be answered without calling it. A pass means the reply comes back with finish_reason: "tool_calls" and the right arguments. When we ran it on GLM 5.3, the call came back as {"city":"São Paulo"}.
MiMo V2.6 Flash passed on 29 September 2026. GLM 5.3 and Qwen3.8 27B passed on 2 October, and DeepSeek V4.1 Flash on 18 October. A model that replies in prose, invents a format, or errors on the tool array does not go in, however good its writing is.
What fails it
On 20 September 2026 we put every other uncensored model in our upstream catalogue through the same test. None of them returned a working tool call. Euryale 70B is the clearest case: its listing says it supports tools, and it returns a provider error the moment a tool array is attached to the request.
That gap between what a listing claims and what the model does is the reason we test instead of reading spec sheets. More on why it happens in why most uncensored models cannot run an agent.
Only the line is served
The gateway serves the five Xploded ids and nothing else. Ask for any other model and the request is refused as an unknown model before it leaves, at no charge. You never end up on an untested model because of a typo or an old config file.
What we change, and what we do not
We do not retrain the weights. Each Xploded model is an uncensored or abliterated build of a current frontier release, served as it is. Our part is the standard and the test, plus everything around the model: an OpenAI-compatible endpoint, an Anthropic-compatible /v1/messages endpoint for Claude Code, and hard spend caps on every key. The line is listed on the models page.