Comparisons

caveira vs running uncensored models locally

You can run an abliterated model yourself, in Ollama, LM Studio or llama.cpp, and drive it with an open-source agent such as OpenCode, Aider or Codex’s CLI. It is free, private and offline, and limited by your hardware. caveira is the hosted path: the agent and large abliterated models with nothing to host, at the cost of an account and sending requests to hosted models.

Checked November 22, 2026 against each product’s own pages, linked below.

At a glance

WhatcaveiraLocal models
PrivacyRequests go through your caveira account to hosted modelsNothing leaves your machine
OfflineNoYes
HardwareAny machine that runs a terminalAbout 19 GB of weights for a 30B model at 4-bit; about 142 GB for a 235B model
Model sizeLarge models with up to a million tokens of contextWhat fits in your memory, usually with a shorter context
SetupInstall, then /loginA runtime, a model download and an agent you configure
CostFree to start. Lightweight $20, Middleweight $50, Heavyweight $100 and Champion $200 a month, with 18x the usage of Freex, 2x the usage of Lightweightx, 5x the usage of Lightweightx and 10x the usage of Lightweightx the usage of Free.Free software (Ollama and llama.cpp are MIT); you pay for the hardware and power

Where Local models is better

  • Total privacy: prompts, code and binaries never leave the machine, which some engagements require.
  • Works air-gapped and offline.
  • No account, no vendor policy and no usage limits beyond your hardware.
  • Free, open-source runtimes and agents, and any model you like.

Where caveira is better

  • Large models without a large GPU: the biggest open models need well over 100 GB of memory to run locally.
  • Up to a million tokens of context, enough for big decompiled binaries and codebases.
  • Nothing to set up: the agent, models, security skills, rea, box and labs work on first run.
  • Frontier models on the same plan, when a stock model is the better tool.

Which to choose

If the work must stay on your machine, or you already own the hardware, run an abliterated model locally and drive it with OpenCode, Aider or Codex’s CLI. Abliteration can cost some model quality, so pick the largest model you can run. If you want large models, long context and the security kit without running any of it, caveira is the faster path.

Questions

What is an abliterated model?

One whose refusal behaviour has been removed by editing its weights. Arditi et al. (2024) showed refusal is mediated by a single direction in a model’s activations; abliteration projects that direction out. It can cost some quality, which further fine-tuning mostly recovers.

What hardware do I need to run an uncensored model locally?

A 30B model at 4-bit quantization is about 19 GB of weights, so a 24 GB GPU or a Mac with 32 GB or more of unified memory, with room left for context. The largest open models are 140 GB and up at 4-bit.

Can caveira use my local model?

No. caveira runs its models through your caveira account; there is no setting for your own endpoint. For a local model, use an open-source agent such as OpenCode or Aider.