Uncensored models, by family.

What each family is

DeepSeek, 6 builds: DeepSeek’s open weights under MIT: V4 Pro (1.6T total, 49B active) and V4.1 Flash (552B backbone, 8B to 16B active), plus the R1 distills of Llama and Qwen. The V4 builds here reason before they answer unless reasoning is switched off. Output from $1.6975 per million tokens on DeepSeek V4.1 Flash Abliterated.

Qwen, 18 builds: Qwen’s Qwen3.8 27B (dense, Apache-2.0, 262,144 tokens natively) and the Qwen3.5 27B line: the derestricted builds with the refusal removed, and the roleplay and fiction tunes the community trains on them. Output from $1.071 per million tokens on Qwen3.5 27B Blossom V6.4 Derestricted.

Gemma, 36 builds: Google’s Gemma 4 line at 12B, 26B-A4B and 31B, the base of most roleplay tunes on this key, plus the Heretic and unshackled builds with the refusal removed from the weights. Output from $0.875 per million tokens on Gemma 4 12B Starry Ocean.

GLM, 11 builds: Z.ai’s GLM-5.3 (753B, MIT), GLM-5.3 Flash (320B total, 18B active) and GLM-4.7 (358B), each served as an uncensored or abliterated build; the Flash builds are the lowest-priced uncensored models on the key. Output from $0.3802 per million tokens on GLM-5.3 Flash Uncensored.

Llama, 4 builds: Meta’s Llama 3 and 3.3 70B with the refusal direction removed, and the Llama 3.1 roleplay tunes. Output from $1.632 per million tokens on Hermes 3 Llama 3.1 70B.

Outside the five families, 61 builds: Roleplay Uncensored, Kimi K3 Uncensored, MiniMax M3 Uncensored, Nemotron 3 Ultra 550B Uncensored, Dolphin Mistral 24B Venice Edition, Hermes 4 405B, Anubis 70B v1.1 and Artemis v1.1 and 53 more. Kimi, MiniMax, Nemotron, MiMo, Hy3, Laguna and Inkling are the official open weights served without a filter; the Mistral tunes are roleplay models; the house builds are served under this site's own ids. All are in the full list below and on the models page.

Download, or call by API

Most of these builds are published on Hugging Face, and the local lists rank them by what fits a card: about 5 GB for an 8B model at Q4, 16 GB for a 27B, 40 GB and more for a 70B. Running the weights yourself means no per-token bill and prompts that never leave the machine, at the price of the hardware and the time to download, update and swap models.

Calling the same build here needs none of that: one base URL, one key and a model id reach every row in the table from a phone, a server or a roleplay client at once, billed per token from a prepaid balance. The 5 family pages show which builds have a card on record and which license it carries, so a download is one click from the API page when prompts must stay local.

Setup mistakes

• SillyTavern and the OpenAI SDK take the base URL ending in /v1. Janitor AI takes the full URL ending in /chat/completions.
• The key goes in the API key field or an Authorization: Bearer header. A key in the URL is refused.
• Model ids are exact and lowercase: glm-5.3-uncensored, not GLM-5.3 Uncensored. GET /v1/models lists every id the key can call.
• A reasoning build with a 300-token reply limit stops mid-thought. Give the DeepSeek and GLM-5.3 builds 1,000 tokens or more, or switch reasoning off; the roleplay tunes answer at once and need no such room.
• A jailbreak preamble is resent and billed on every message. None of these builds needs one, and a refusal written as ordinary text is billed like any answer.

Uncensored models: questions