Does OpenRouter have uncensored or abliterated models?

Silk ComputeComparison2 min read

Diagram: a router splits into six gray paths to models with refusals intact, and an orange path runs from Silk Compute to a model with refusal behavior removed

OpenRouter's public model list describes none of its 451 text models as abliterated as of November 2, 2026, and only two as uncensored, both 24B fine-tunes. For answers without refusals on a large model, use an API built for it, like Silk Compute.

Does OpenRouter have abliterated models?

Not by its own descriptions. We searched the id, name and description of every model in its public list. None says abliterated, and these are the only two that say uncensored, with our models for comparison.

What it isContext
Venice: Uncensored, on OpenRouterDolphin Mistral 24B fine-tune128K tokens
TheDrummer: Cydonia 24B V4.1, on OpenRouterCreative writing model based on Mistral Small 3.2 24B131K tokens
Silk ComputeAbliterated DeepSeek V4.1 Flash 1017, GLM-5.3 and GLM-5.3 Flash1M tokens

From OpenRouter's public model list as of November 2, 2026.

Routing does not remove refusals

A router is a switchboard, not a surgeon. It passes your request to the model you pick. If that model was trained to refuse, it still refuses. To get answers without refusals, the model itself has to be uncensored, either fine-tuned or abliterated.

What should you check in an uncensored model?

Once you're shopping for the model itself, look for:

  • Size. Many uncensored models are small fine-tunes. They're fine for chat, but weaker at long coding and research tasks.
  • Context and output limits. Agents read a lot and write a lot. Short limits end runs early.
  • Tool calling. A coding agent is stuck without it, so make sure it works before you plan around a model.

What does Silk Compute do differently?

We serve only uncensored models, and only large ones: Kimi K3.1 Pro, Hy4, DeepSeek V4.1 Flash 1017, GLM-5.3 and GLM-5.3 Flash, abliterated, with 1M tokens of context each. Every model speaks the OpenAI, Anthropic Messages and OpenAI Responses formats, so Claude Code, Codex and OpenCode work out of the box, on plans from $10 a month.

You don't have to leave OpenRouter

Keep OpenRouter for mainstream models and add Silk Compute for the work they refuse. Both are OpenAI compatible: change the base URL to https://api.silkcompute.com and the model id, and the rest of the request keeps its shape. The quickstart takes about two minutes. Comparing dedicated APIs? See our abliteration.ai and Venice alternatives.

More from the blog