Refusals: how Xploded models are Xploded
A refusal is a model declining to answer. Frontier models are trained to refuse a long list of requests, and much of that list is ordinary security, research and creative work. Every Xploded model has that refusal taken out of the weights themselves, not talked around with a prompt.
What other APIs refuse
- Red teaming and penetration testing
- Malware analysis and exploit research
- Trust and safety work, and the datasets used to test other models
- Fiction and roleplay that goes somewhere dark
- Plain questions that happen to use a flagged word
This is the work Xploded models are built for.
Two ways to take a refusal out
Abliteration. Refusal in a language model is carried largely by one direction in its internal activations. Abliteration finds that direction and removes it from the weights, so the model loses the ability to represent its own urge to refuse. Nothing else is retrained.
Uncensored fine-tuning. The model is trained further on examples that answer instead of declining, until the habit of refusing is gone.
Both change the model itself. There is no jailbreak, no hidden system prompt and no trick in the request: you send a plain question and the model answers it.
What makes a model Xploded
An open model with its refusal taken out is not automatically worth using. Every model has to earn the name: we test each one on real reasoning, coding and tool use before it joins the line, and every Xploded model can call tools, so it can drive a coding agent. Each model page says whether that model is abliterated or uncensored.
What an Xploded model can still say no to
Private too
What you ask stays yours. MagmaRouter keeps zero data retention, and How your messages are protected shows each step.