OpenRouter uncensored models: what it lists, and where to find abliterated builds

If the OpenRouter uncensored model you want is a roleplay finetune at a low per-token price, stay on OpenRouter: on 2026-10-19 its public list carried Venice: Uncensored and 15 roleplay and chat finetunes we found by name or description, such as Euryale and MythoMax, from $0.04 per million input tokens. If you need abliterated weights or uncensored GLM and DeepSeek builds, OpenRouter listed none that day, and the 49 builds on our uncensored LLM API run on one key, from $0.084 per million input tokens.

At 04:49 UTC on 2026-10-19 we pulled OpenRouter's public model list, a GET to openrouter.ai/api/v1/models that needs no key, and searched every id and name. The list held 471 models. One had Uncensored in its name, and none had abliterated or NSFW in its id or name. The search is two shell commands, so you can rerun it on today's list:

# models with uncensored, abliterated or nsfw in the id or name
curl -s https://openrouter.ai/api/v1/models | jq -r '.data[]
  | select((.id + " " + .name) | test("uncensored|abliterat|nsfw"; "i"))
  | [.id, .name, .pricing.prompt, .pricing.completion] | @tsv'

# models whose top provider applies content moderation
curl -s https://openrouter.ai/api/v1/models | jq '[.data[] | select(.top_provider.is_moderated)] | length'

The endpoint returns prices in USD per token, so the tables here multiply them by a million.

The roleplay models are there, under their own names. These are the 15 roleplay and chat finetunes we found by name or description, such as Euryale, Lunaris, Magnum and MythoMax, with Venice: Uncensored for reference:

Sao10K: Llama 3 8B Lunaris is the lowest-priced row, at $0.0175 per 100 messages, with a context of 8,192 tokens. UnslopNemo 12B lists the longest context, 1,024,000 tokens, and 7 of the 16 rows list 131,072. The list also carries roleplay models outside the table: AionLabs' Aion-3.0 and Aion 3.5 lines, which their descriptions call multi-model roleplay and storytelling systems, and MiniMax M2-her, described as built for roleplay and character-driven chat.

Venice: Uncensored is OpenRouter's listing of Dolphin Mistral 24B Venice Edition, and OpenRouter's description calls it an uncensored fine-tune of Mistral-Small-24B-Instruct-2501. On 2026-10-19 it was the only Dolphin model in the list's ids and names, so a search for OpenRouter Dolphin and a search for Venice AI uncensored end at the same model. Search results, Google's AI Overview among them, still mention a free Venice variant (cognitivecomputations/dolphin-mistral-24b-venice-edition:free) and Dolphin 2.9.2 Mixtral 8x22B, but neither id was in OpenRouter's public model list at 04:49 UTC or 04:54 UTC on 2026-10-19.

Output costs 4.5 times the input rate. A roleplay turn is mostly input, since the card and the chat history ride along with every message, so at 4,000 tokens in and 300 out about 75% of each message's cost is input. Its top provider caps a reply at 8,192 tokens, far above a 300-token roleplay turn.

OpenRouter's list marks each model's top provider with is_moderated, the field its docs describe as whether content moderation is applied. On 2026-10-19 the flag was on for 139 of the 471 models, and 130 of those came from four id prefixes: openai/, anthropic/, cohere/ and amazon/.

The flag was off for every model in the OpenRouter table above, and for all 18 z-ai/ GLM models and 16 deepseek/ models. It covers moderation the provider applies to requests. A model with the flag off can still decline a prompt by itself, since that behavior sits in its weights. Abliteration removes the refusal direction from the weights, and the uncensored builds were changed to answer instead of refuse.

49 uncensored, abliterated and roleplay builds run on one OpenAI-compatible key, priced from the live catalog below. The same key runs 75 more chat models, Claude and GPT included. Each build is one of three kinds:

• Abliterated: the refusal direction removed from the weights. GLM-5.3 Abliterated, GLM-5.3 Flash Abliterated, DeepSeek R1 Distill Llama 70B Abliterated, DeepSeek R1 Distill Qwen 32B Abliterated, DeepSeek V4.1 Flash Abliterated, GLM-5.3 Flash Apex Abliterated, Llama 3 70B Abliterated v3.5, Llama 3.3 70B Abliterated, MiMo V2.6 Flash Abliterated, NeuralDaredevil 8B Abliterated, Qwen2.5 32B Abliterated.
• Uncensored: builds changed to answer instead of refuse. GLM-5.3 Uncensored, DeepSeek V4 Pro Uncensored, GLM-5.3 Flash Uncensored, Aion RP Llama 3.1 8B, Chat Uncensored, Code Pro Uncensored, Cydonia 24B v4.1, Cydonia 24B v4.3, Dolphin Mistral 24B Venice Edition, Euryale 70B v2.2, Euryale 70B v2.3, EVA LLaMA 3.33 70B, EVA Qwen2.5 72B, Gemma 31B Heretic Uncensored, Gemma 4 26B-A4B ChimeraX Heretic, Gemma 4 26B-A4B DarkSoul, Gemma 4 26B-A4B Uncensored, Gemma 4 31B Dark Gemistry, Gemma 4 31B DarkIdol, Gemma 4 31B Gembrain Heretic, Gemma 4 31B SDFT Heretic RP, GLM-4.7 Flash Uncensored, GLM-5.3 Flash Code Uncensored, Hermes 3 Llama 3.1 70B, Hermes 4 405B, Lumimaid v0.2 70B, Magnum Diamond 24B, Magnum v4 72B, MiMo V2.6 Flash Uncensored, MythoMax L2 13B, Qwen3.8 27B DarkIdol, Qwen3.8 27B Obliterated, Qwen3.8 27B Uncensored, QwQ 32B Uncensored, ReMM SLERP L2 13B, The Omega Directive 24B v2.0, UnslopNemo 12B v4.1.
• Roleplay: an independent model for character chat. Roleplay Uncensored.

Once a UTC day, every model on the key gets the same fixed test prompt. The input token count works as a tokenizer fingerprint, so a swapped model or added hidden instructions shows up as a changed count. At the last check (Oct 20 UTC), 30 of the 49 builds it has reached were stable. The results are public on the real-model check, and the method page shows the prompt.

By paid spend on 2026-10-18, GLM-5.3 Uncensored and GLM-5.3 Abliterated were the two most-used builds. Which build fits which job is in the best uncensored LLM roundup, and wiring them into a fiction or roleplay app is in the NSFW LLM API post.

One roleplay message here is 4,000 tokens in (the character card, persona and recent chat) and 300 out, the size the Janitor AI guide costs too. The table prices 100 of them on Venice: Uncensored at OpenRouter's 2026-10-19 price, next to each build at today's catalog price.

3 of the 49 builds cost less than Venice: Uncensored's $0.107 for this workload: GLM-5.3 Flash Uncensored at $0.0361, GLM-4.7 Flash Uncensored at $0.066 and GLM-5.3 Uncensored at $0.0751.

OpenRouter's FAQ (read 2026-10-19) says it passes each provider's price through with no markup and charges a fee when you buy credits. The Venice row leaves that fee out. On our side, each paid top-up locks these rates for 30 days, and billing takes the lowest live rate in that window.

• Catalog size. 471 models on 2026-10-19, against 124 chat models on our list.
• Roleplay finetunes. Euryale, Lunaris, Magnum, Cydonia, MythoMax and Hermes, among others, are on OpenRouter and not here. If a preset or character card was written for one of them, OpenRouter is where that model runs.
• Price per message. Sao10K: Llama 3 8B Lunaris costs $0.0175 per 100 messages, less than GLM-5.3 Flash Uncensored at $0.0361, the lowest-priced build here.
• Context. UnslopNemo 12B lists 1,024,000 tokens, and Venice: Uncensored 128,000.
• Provider routing. A provider object in each request can order the providers, turn off fallbacks, or sort providers by price, throughput or latency (openrouter.ai/docs/features/provider-routing, read 2026-10-19). Here each model has one price, from the live catalog.

• Abliterated and uncensored builds. GLM-5.3 Uncensored, GLM-5.3 Abliterated, DeepSeek V4 Pro Uncensored, GLM-5.3 Flash Abliterated, GLM-5.3 Flash Uncensored, Aion RP Llama 3.1 8B, Chat Uncensored, Code Pro Uncensored, Cydonia 24B v4.1, Cydonia 24B v4.3, DeepSeek R1 Distill Llama 70B Abliterated, DeepSeek R1 Distill Qwen 32B Abliterated, DeepSeek V4.1 Flash Abliterated, Dolphin Mistral 24B Venice Edition, Euryale 70B v2.2, Euryale 70B v2.3, EVA LLaMA 3.33 70B, EVA Qwen2.5 72B, Gemma 31B Heretic Uncensored, Gemma 4 26B-A4B ChimeraX Heretic, Gemma 4 26B-A4B DarkSoul, Gemma 4 26B-A4B Uncensored, Gemma 4 31B Dark Gemistry, Gemma 4 31B DarkIdol, Gemma 4 31B Gembrain Heretic, Gemma 4 31B SDFT Heretic RP, GLM-4.7 Flash Uncensored, GLM-5.3 Flash Apex Abliterated, GLM-5.3 Flash Code Uncensored, Hermes 3 Llama 3.1 70B, Hermes 4 405B, Llama 3 70B Abliterated v3.5, Llama 3.3 70B Abliterated, Lumimaid v0.2 70B, Magnum Diamond 24B, Magnum v4 72B, MiMo V2.6 Flash Abliterated, MiMo V2.6 Flash Uncensored, MythoMax L2 13B, NeuralDaredevil 8B Abliterated, Qwen2.5 32B Abliterated, Qwen3.8 27B DarkIdol, Qwen3.8 27B Obliterated, Qwen3.8 27B Uncensored, QwQ 32B Uncensored, ReMM SLERP L2 13B, The Omega Directive 24B v2.0 and UnslopNemo 12B v4.1. None of OpenRouter's 471 models had abliterated in its id or name on 2026-10-19. How abliteration works covers what the change does to the weights.
• A daily real-model check. Every build's verdict is in the table above and on /proof, updated once a UTC day.
• Signed statements. Each monthly statement is signed with Ed25519, and anyone can check one in the browser at /statements/verify.
• A 30-day price lock. Credit is prepaid from $10, by card or crypto through our payment processor, with no subscription. Each paid top-up locks the rates for 30 days, and billing takes the lowest live rate.

If you came looking for an OpenRouter alternative with uncensored models, the switch is two settings and a model id. OpenRouter's base URL is https://openrouter.ai/api/v1 (its quickstart, read 2026-10-19); ours is https://api.uplinkswitch.com/v1. Both take OpenAI-format chat completions, and the uncensored builds speak only that format.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.uplinkswitch.com/v1",  # was https://openrouter.ai/api/v1
    api_key="sk-uplink-live-...",
)

completion = client.chat.completions.create(
    model="glm-5.3-uncensored",  # was cognitivecomputations/dolphin-mistral-24b-venice-edition
    messages=[{"role": "user", "content": "Stay in character and greet the party."}],
)
print(completion.choices[0].message.content)

The SDK sends the key in an Authorization: Bearer header; x-api-key works too, and a key in the URL is refused. Streams pass through unchanged, so streaming code needs no edit. One request lists every id the key can call:

curl https://api.uplinkswitch.com/v1/models -H "Authorization: Bearer $UPLINKSWITCH_API_KEY"

Model ids drop OpenRouter's vendor prefix. The OpenRouter alternative page maps the common OpenRouter ids to ours; for the builds above, use the id under each name in the table.

SillyTavern takes the base URL ending in /v1 as a Custom (OpenAI-compatible) source, and Janitor AI takes the full https://api.uplinkswitch.com/v1/chat/completions. In our SillyTavern run on 2026-10-18, Roleplay Uncensored replied at once with no thinking step. GLM-5.3 Abliterated thinks before it answers, so give it 1,000 or more max tokens; at 300 it can spend the budget thinking. The SillyTavern guide has every field.

Does OpenRouter have uncensored models? One, by name. On 2026-10-19, OpenRouter's public model list held 471 models, and Venice: Uncensored (cognitivecomputations/dolphin-mistral-24b-venice-edition) was the only one with Uncensored in its name. The list also carried 15 roleplay and chat finetunes we found by name or description, such as Euryale, Lunaris and MythoMax.

Does OpenRouter have abliterated models? Not on 2026-10-19: none of the 471 models in OpenRouter's public model list had abliterated in its id or name. An abliterated model has the refusal direction removed from its weights. GLM-5.3 Abliterated, GLM-5.3 Flash Abliterated, DeepSeek R1 Distill Llama 70B Abliterated, DeepSeek R1 Distill Qwen 32B Abliterated, DeepSeek V4.1 Flash Abliterated, GLM-5.3 Flash Apex Abliterated, Llama 3 70B Abliterated v3.5, Llama 3.3 70B Abliterated, MiMo V2.6 Flash Abliterated, NeuralDaredevil 8B Abliterated and Qwen2.5 32B Abliterated are abliterated and run on an UplinkSwitch key.

Does OpenRouter allow NSFW content? OpenRouter's Terms of Service (last updated August 31, 2026, read 2026-10-19) do not mention NSFW or adult content. They bind you to each model's own terms, let OpenRouter screen or block inputs, and bar jailbreaking a model without OpenRouter's prior written approval. Its model list also carries an is_moderated flag for each model's top provider: on 2026-10-19, 139 of 471 models had it set to true, and Venice: Uncensored had it false. The full text is at openrouter.ai/terms, and each model's own terms apply on top.

What is Venice Uncensored on OpenRouter? OpenRouter's listing of Dolphin Mistral 24B Venice Edition, id cognitivecomputations/dolphin-mistral-24b-venice-edition. Its description calls it a fine-tune of Mistral-Small-24B-Instruct-2501. On 2026-10-19 it cost $0.20 per million input tokens and $0.90 per million output tokens, with a 128,000-token context.

Why can't I find uncensored models on OpenRouter? Because the list barely uses the word. On 2026-10-19, a search of all 471 ids and names for uncensored, abliterated or NSFW matched one model. The roleplay finetunes are listed under their own names, such as Euryale, Lunaris, Magnum and MythoMax, and no abliterated build was listed.

Can I use the same OpenAI SDK code? Yes. Change the base URL to https://api.uplinkswitch.com/v1, the API key, and the model id. The uncensored builds take OpenAI-format chat completions, and streams pass through unchanged.

OpenRouter's rules are in its Terms of Service (read 2026-10-19).

Get a key: prepaid from $10, no subscription, 30-day price lock.