Brutal AI

Brutal AIModelsGuidesUncensored LLM API providers compared (2026)

Comparison7 min read

Uncensored LLM API providers compared (2026)

Seven APIs people use when stock models refuse: how each one actually removes refusals, which formats it speaks, whether it works with Claude Code, what it logs and where it runs. Every fact links to the provider's own page, checked on 17 November 2026.

Short answer: for abliterated models with an OpenAI- and Anthropic-compatible API, the main options are XField by Brutal AI, abliteration.ai and Venice. Featherless hosts many community abliterated models on an OpenAI-compatible API. OpenRouter and NanoGPT are routers: they pass your request to a model hosted by someone else. Together AI hosts mainstream open models but lists no uncensored ones. If you need EU hosting, XField is hosted in the EU and abliteration.ai offers EU residency as an option; the others do not state a region or are US-based.

We run XField, so read this with that in mind. We have kept the table to facts each provider publishes, linked every one, and written "not stated" where a provider does not say.

Three ways providers remove refusals

  • Abliterated weights. The direction in the model that leads to refusals is removed from the weights. Knowledge and skills stay; false refusals go. See How abliteration works.
  • Uncensored fine-tunes. A community model such as Dolphin or Hermes is trained further to refuse less. Quality depends on the fine-tune and usually trails the newest base models.
  • Routing. A router forwards your request to another provider's model. It adds one key and one bill, but it cannot change what the weights do. See Routing doesn't remove refusals.

The comparison

ProviderApproachOpenAI APIAnthropic Messages / Claude CodePrompt loggingHosting regionAccess
XField (Brutal AI)Own abliterated versions of DeepSeek V4.1 Flash, GLM 5.3, Kimi K3.1 Pro, Qwen 3.8 27BYes: Chat Completions and ResponsesYes, native; Claude Code documentedNot stored, not used for trainingEUOn request, reviewed per company
abliteration.aiOwn abliterated models; optional content policies per projectYes: Chat Completions and ResponsesMessages endpoint yes; Claude Code documentedProcessed in memory, not written to disk; no trainingUS by default, EU optionalSelf-serve, pay as you go or plans
VeniceOwn uncensored and abliterated variants; also resells other providers' models, some in an anonymized modeYesNo native endpoint; Claude Code via a proxyNot stored for Venice-hosted models; training not statedNot statedSelf-serve: API key, or USDC pay per request
FeatherlessHosts community abliterated and uncensored models from Hugging FaceYesNot statedNot logged; training not statedNot statedMonthly plans
OpenRouterRouter; its uncensored model is Venice Uncensored, served by VeniceYesMessages yes; Claude Code only guaranteed with Anthropic's own modelsOnly if you opt in; each upstream provider has its own policyEU routing on Business and Enterprise plansSelf-serve credits
NanoGPTRouter; lists abliterated models modified to refuse lessYesYes, names Claude CodeNot logged in normal inference; some routes allow trainingNot stated; EU-only processing per requestSelf-serve balance or subscription
Together AIMainstream open models; no uncensored models listed; terms restrict contentYesMessages endpoint yesNot stored under zero data retentionUS endpoint availableSelf-serve pay per token

Example list prices per 1M tokens (input / output), seen 17 Nov 2026: abliteration.ai Abliterated Large v2 $3.00 / $5.00; Venice Uncensored 1.2 $0.20 / $0.90; OpenRouter Venice: Uncensored $0.20 / $0.90. Prices change often, so check the provider's page. XField pricing comes with access.

Who serves Kimi K3.1 Pro uncensored?

Kimi K3.1 Pro is Moonshot AI's flagship open model. Most providers serve only the stock version, which still refuses. Checked on 17 November 2026:

ProviderKimi K3.1 Pro
XField (Brutal AI)Abliterated, our own: kimi-k3.1-pro-uncensored, 1M context, up to 922K output, video input, hosted in the EU
NanoGPTAn abliterated version listed from a third-party host (Lyber), next to the stock model
VeniceStock model only
OpenRouterStock model only, from several providers
FeatherlessStock model only
Together AIStock model only
abliteration.aiNo Kimi model listed

Sources: NanoGPT models, Venice models, OpenRouter, Featherless models, Together serverless models, abliteration.ai pricing.

How to choose

  • You run a coding agent such as Claude Code. Pick an API with a native Anthropic Messages endpoint and large models with working tool calls: XField, abliteration.ai, NanoGPT or Together (Together has no uncensored models). XField and abliteration.ai document Claude Code setups. Setup for XField: Use Claude Code with an uncensored model.
  • You need EU hosting or data residency. XField is hosted in the EU. abliteration.ai offers EU residency as an option, OpenRouter on its Business and Enterprise plans.
  • You want the newest large open models, abliterated. XField serves abliterated DeepSeek V4.1 Flash, GLM 5.3 and Kimi K3.1 Pro. Venice and abliteration.ai also serve abliterated versions of recent models.
  • You want to try many small community models cheaply. Featherless has the widest catalog of community abliterated models.
  • You want one key for everything. A router like OpenRouter or NanoGPT. Check which upstream host serves the uncensored model and its logging policy.

What to check before you send sensitive data

  1. Who hosts the weights. With a router, your prompt goes to another company. Its policy applies.
  2. Whether prompts are stored or used for training. Read the provider's data page, not the marketing page.
  3. Where the servers are. Most providers do not state a region.
  4. Whether the model actually answers. Run five prompts from your real work, the ones your current model refuses, before committing.

One-to-one comparisons

Setup guides for your tools: Claude Code, Codex CLI, OpenCode, Cline, Roo Code, Continue, Aider.

About XField

XField is the model API of Brutal AI in Vienna. We built it for our own engineers, who use these models on client projects for security audits and red-teaming, and we open spare capacity to a few teams. It is hosted in the EU, speaks OpenAI Chat Completions, Responses and Anthropic Messages, and does not store prompts or completions (data handling). Access is reviewed company by company under our acceptable use policy. Request a key.

Sources (checked 17 Nov 2026)

Access

Access is by invitation and reviewed per company. Tell Lukas what you are building and where your current model refuses. Pricing comes with the invitation.

lukas@brutal.ai

Uncensored models don't change what is allowed: you stay responsible for use and results under our acceptable use policy.

More guides