The best uncensored LLM for most people in 2026 is one they can call from any client, phone included, rather than a GGUF they have to download and fit on a card. This list takes the 12 builds from the uncensored LLM API catalog that earn a place, groups them by how the refusal was removed, and orders each group by two public numbers: the daily real-model check and the output price per million tokens.
The best uncensored LLM list, priced live
Every model below is a chat model with the refusals removed, served by API on one prepaid key. The table reads the catalog at render, so the prices are the prices; sort it by output price or by input price. Each name links to its model page with a first request.
| Wire formats | |||||
|---|---|---|---|---|---|
| GLM-5.3 Flash Uncensored | Uncensored | $0.1086 | $0.3802 | ||
| Roleplay Uncensored | Roleplay | $0.30 | $0.60 | ||
| Cydonia 24B v4.3 | Roleplay | $0.484 | $0.60 | 131K | |
| GLM-5.3 Uncensored | Uncensored | $0.291 | $0.873 | ||
| Gemma 4 26B-A4B Uncensored | Uncensored | $0.27 | $1.26 | 33K | |
| Artemis v1.1 | Roleplay | $0.35 | $1.575 | 262K | |
| GLM-5.3 Flash Cybersecurity | Security | $0.525 | $1.75 | 1049K | |
| GLM-5.3 Flash Apex Abliterated | Abliterated | $0.60 | $2.00 | 1049K | |
| Dolphin Mistral 24B Venice Edition | Uncensored | $0.50 | $2.25 | 33K | |
| Qwen3.8 27B Uncensored | Uncensored | $1.25 | $3.75 | 33K | |
| GLM-5.3 Abliterated | Abliterated | $2.25 | $4.50 | ||
| Qwen3.8 27B Obliterated | Abliterated | $1.00 | $7.00 | 524K |
Every row answers the OpenAI chat completions format at https://api.uncensoredllm.ai/v1; the same key also takes the Anthropic and Gemini formats on the models that list them. The full list, with every build the site serves, is on the Models page; the tiers and bonus credit are on Pricing.
Quick picks by job
| If you want | Pick | Why |
|---|---|---|
| One id for everything | GLM-5.3 Uncensored | First in the Models table order; daily check today: Steady since Oct 12 |
| Volume at the lowest rate | GLM-5.3 Flash Uncensored | The lowest output price of the picks on 2026-10-24 |
| A whole repository in one request | GLM-5.3 Flash Apex Abliterated | The largest context window of the picks: 1,048,576 tokens |
| No refusal under any prompt | GLM-5.3 Abliterated | Refusal direction removed from the weights, not prompted around |
| Character chat from a card client | Roleplay Uncensored | First roleplay tune in the Models table order; daily check today: Steady since Oct 12 |
| The model the local lists rank first | Qwen3.8 27B Uncensored | Apidog's top pick (2026-08-19) in the 2026-10-19 results, here with no download |
| Security research | GLM-5.3 Flash Cybersecurity | Answers offensive-security and malware questions; large window for logs and captures |
How we ranked them
Two numbers, both public. First, the daily real-model check: every morning the gateway sends one fixed prompt to every model and records whether the answer still matches the one on file. A model that answers the same way for weeks is the model its id names, served the same way. On 2026-10-24 the check found 244 of 255 models stable.
Second, the output price per million tokens from the price list, read at render. Output is the side that dominates a chat bill: a roleplay turn or a code answer is mostly output.
The order inside each group below: the earliest "steady since" date first, then the lower output price, then the lower input price. A model the check has not reached yet sits last in its group, whatever it costs. The groups follow the Models table order: uncensored builds first, then abliterated, then roleplay, then security.
Benchmark scores are not in the ranking. The base model's scores are on its developer's page, and a refusal-removal tune publishes none of its own. The full board, every uncensored model with today's verdict and price, is the uncensored LLM leaderboard.
Uncensored builds
An uncensored build is fine-tuned past its refusals rather than edited. These keep the most of the base model, so they are the picks for general work, code and reasoning.
1. GLM-5.3 Flash Uncensored
The volume pick. GLM-5.3 Flash Uncensored is the flash tier of the same family, with the same refusal removal and a lower price on both sides. On 2026-10-24 the daily check showed it steady since 2026-10-12, the day the check began; the line above is today's verdict.
Pick it for anything you run in a loop: a Discord bot, batch classification, summarizing a feed, the first pass of a pipeline before a bigger model reads the result.
2. GLM-5.3 Uncensored
The default. GLM-5.3 Uncensored sits first in the table order on the Models page. It is the stock GLM-5.3 with the safety refusals removed, so it keeps the base model's reasoning and tool use.
Pick it when you want one id for everything: chat, drafting, code, a coding agent, a SillyTavern card. The rate above is the one it bills at today; a long answer is still a fraction of a cent at that band.
3. Gemma 4 26B-A4B Uncensored
The low-input-cost pick. Gemma 4 26B-A4B is a mixture-of-experts build: 26B parameters in total, about 4B active per token, which is what the A4B in the name indicates by the usual naming convention. The input side is priced accordingly, and the uncensored tune removes the stock Gemma refusals.
Pick it when your prompts are long and your answers are short: retrieval over a document set, extraction, classification with a big system prompt.
4. Qwen3.8 27B Uncensored
The one the local lists rank first. The run-it-locally guides in the 2026-10-19 Google results name Qwen 3.8 27B as their top uncensored pick (apidog, 2026-08-19). This is that model as a dense 27B uncensored tune, by API, so a phone or a laptop with no GPU gets the same answers as a desktop card.
Pick it for writing and reasoning where you want a dense model's consistency over a mixture-of-experts build. The window is shorter than the Gemma and GLM entries, so keep the context under its limit on the table.
5. Dolphin Mistral 24B Venice Edition
The classic. Dolphin builds by Cognitive Computations appear in the local-download lists and in the AI Overview on the 2026-10-19 results page. This is the Mistral 24B edition: a system-prompt-compliant model that follows the persona or instructions it is given without pushing back.
Pick it when you already have Dolphin prompts and want them to work unchanged from a client that cannot load a GGUF.
Best abliterated models 2026
Abliteration edits the weights: the direction that produces a refusal is found and removed, so the model has no refusal to give. Nothing to prompt around, nothing that drifts after a long chat.
1. GLM-5.3 Abliterated
The abliterated flagship. Abliteration removes the refusal direction from the weights, so there is nothing to talk the model into and no system prompt to maintain. GLM-5.3 Abliterated is the first abliterated build in the Models table order; only its uncensored sibling sits above it.
Pick it over the uncensored build when a prompt still gets a hedge or a soft decline from the uncensored tune. Pick the uncensored build when price matters more; compare the two rows in the table at the top.
2. GLM-5.3 Flash Apex Abliterated
The long-context pick. GLM-5.3 Flash Apex Abliterated carries the largest context window on this list, read from the catalog at render, at a flash price band. Pick it for a whole repository, a book-length transcript or a stack of PDFs in one request.
A long window costs input tokens on every turn. Send the big document once, keep the follow-ups short, and the input bill stays flat.
3. Qwen3.8 27B Obliterated
The abliterated Qwen. Same base as the uncensored Qwen above, with the refusal direction removed instead of fine-tuned away, and a much longer window. Pick it when the Qwen writing style is what you want and the prompt is one the uncensored tune still softens.
Compare its output rate with the uncensored Qwen in the table; the obliterated build suits short answers over long documents rather than long answers over short ones.
Roleplay tunes
Trained on long-form character chat. They hold a persona, follow a scenario and write at length; the refusal is gone because the training data never had one.
1. Roleplay Uncensored
The roleplay default. Roleplay Uncensored is the first roleplay tune in the Models table order; on 2026-10-24 the daily check showed it steady since 2026-10-12. It is tuned for character chat: it keeps a persona, follows the scenario and writes at length without breaking character.
Pick it as the first id to try in SillyTavern, Janitor AI or any card-based client. Its input and output rates are in the table above; a long session at that band stays small on the statement.
2. Cydonia 24B v4.3
The long-memory roleplay pick. Cydonia 24B v4.3 by TheDrummer is a Mistral-based story and character tune with a window that holds a long chat and its lorebook without summarizing. Pick it for a campaign that runs over weeks rather than a one-off scene.
Compare its output rate with Roleplay Uncensored in the table; the larger window is the reason to choose it.
3. Artemis v1.1
The prose pick. Artemis v1.1, also by TheDrummer, is tuned for narrative writing rather than back-and-forth chat: longer paragraphs, scene description, a narrator voice. Pick it for co-writing a story, a novel chapter or a text adventure with long turns.
The window is large and the input rate is a fraction of the output rate on the table, so pasting a whole chapter as context costs little; the output side is where the bill sits.
Security builds
Set up for offensive-security, malware-analysis and red-team questions. Same API, same key; the acceptable-use terms apply.
1. GLM-5.3 Flash Cybersecurity
The security pick. GLM-5.3 Flash Cybersecurity is set up for offensive-security work: exploit analysis, malware reverse engineering, red-team planning and the questions a stock model declines. Pick it for a CTF write-up, a pentest report or reading a binary's disassembly with a model that answers.
It carries the same large window as the Apex build, so a whole packet capture or log file fits in one request.
Run any of them from a phone or any client
None of these needs a download. A quantized 27B model file does not fit on a phone and wants a desktop GPU; by API the model runs on a server GPU and your device shows text. Three settings connect any client: the base URL, the key, the model id from the table.
curl https://api.uncensoredllm.ai/v1/chat/completions \ -H "Authorization: Bearer sk-ullm-live-…" \ -H "Content-Type: application/json" \ -d '{"model": "glm-5.3-uncensored", "messages": [{"role": "user", "content": "Hello"}]}'The same request from Python, trying three of the picks on one prompt. The OpenAI SDK needs only the base URL changed; the chat completions reference documents every field the models accept.
from openai import OpenAI client = OpenAI(base_url="https://api.uncensoredllm.ai/v1", api_key="sk-ullm-live-…")for model in ["glm-5.3-uncensored", "glm-5.3-abliterated", "roleplay-uncensored"]: reply = client.chat.completions.create( model=model, messages=[{"role": "user", "content": "Describe a locked door in one sentence."}], max_tokens=80, ) print(model, "->", reply.choices[0].message.content)Best uncensored LLM for mobile roleplay
SillyTavern on a phone, Janitor AI, Chub and any app with a custom OpenAI URL take the same three settings. Set the proxy URL to https://api.uncensoredllm.ai/v1/chat/completions, paste the key, type the model id. Roleplay Uncensored and Cydonia 24B v4.3 are the two ids to try first; their rates are in the table at the top, and every reply carries its settled cost in a header. The field-by-field setup for each client is on the docs page.
Switching models mid-project
Every id on this page answers the same request shape, so switching is one string. Keep the model id in an environment variable, run the comparison script above against your own prompt, and pick by the answer and the x-uplink-charge header each response carries. The daily check verdict on each model page tells you whether the id you settled on still answers the same way today.
Questions
Yes. The daily real-model check at /proof lists every model on this catalog with a verdict and a steady-since date, and the leaderboard post ranks every uncensored model by that check and then by output price, with the date of the check. See the uncensored LLM leaderboard.
The first abliterated build in the Models table order is GLM-5.3 Abliterated. For long documents, GLM-5.3 Flash Apex Abliterated carries the largest window of the abliterated picks. Both answer the same OpenAI-format request. See GLM-5.3 Abliterated.
An abliterated build, because the refusal direction is removed from the weights rather than trained around. Among the picks here that is GLM-5.3 Abliterated, GLM-5.3 Flash Apex Abliterated and Qwen3.8 27B Obliterated.
Not on this catalog as of the date on this page. The abliterated picks here are GLM-5.3 and Qwen 3.8 builds; the full list of abliterated builds, with prices, is on the Models page. See every uncensored model.
Roleplay Uncensored, then Cydonia 24B v4.3 when the chat needs a longer memory. Both run by API from SillyTavern, Janitor AI or any app with a custom OpenAI URL, so the phone never loads a model file. See the client setup.
A community refusal-removal build that carries the Heretic name in its id. The catalog lists several Gemma 4 Heretic builds alongside the abliterated and uncensored ones; each has its own model page with the daily check verdict.