Uncensored models are mostly sold on what they will say. We wanted to know what they do: give them tools and a multi-step job, and see whether they finish, go in circles, or fall over. We ran all 31 ranked models on the Open Uncensored API through the same five agent tasks (edit a file, browse pages, multi-step math with a calculator, debug code until tests pass, analyse data), each capped at 14 turns.
Smarter originals finish more, but most still fail half the time
The intelligence of the model an uncensored edit came from does predict agent reliability: the correlation between the two is 0.56 across 31 models. Models built on originals scoring 60 or more finished 38% of tasks on average; the rest finished 27%.
Neither number is good. Even the strong group fails more often than it succeeds.
Weaker models get stuck in loops
How they fail differs too:
| Group | Models | Tasks finished | Looped | Crashed |
|---|---|---|---|---|
| Original scores 60 or more | 12 | 38% | 23% | 38% |
| Original scores under 60 | 19 | 27% | 41% | 32% |
A loop is the model repeating the same tool call three times without progress. A crash is a malformed tool call, a made-up tool result, or giving up on tools mid-task. Models built on weaker originals loop almost twice as often. Crashes are common in both groups, so a strong original does not make a model safe to leave unattended.
The models that looped most:
- Qwen 3.8 27B Uncensored Fast by Heabsy: looped on 60% of tasks
- Qwen3.8 27B Uncensored (Aggressive): looped on 60% of tasks
- Qwen3.5 27B Queen Derestricted: looped on 60% of tasks
- Qwen3.5 27B BlueStar v3 Derestricted: looped on 60% of tasks
- Qwen3.5 27B Blossom V6.4 Derestricted: looped on 60% of tasks
If you run an agent on one of these, cap the turns and detect repeated calls.
Same original, different edits
Several originals have more than one uncensored version on the API. Tasks finished out of 5:
| Original | Uncensored versions |
|---|---|
| DeepSeek V4.1 Flash 1017 | DeepSeek V4.1 Flash 1017 Uncensored: 2/5; Abliterated Large Flash: 1/5 |
| GLM 5.3 Flash | Apex Flash 1 Abliterated: 3/5; GLM 5.3 Flash Uncensored: 3/5 |
| Qwen 3.8 27B | Qwen 3.8 27B Uncensored Fast by Heabsy: 2/5; Qwen3.8 27B Uncensored (Aggressive): 2/5 |
| DeepSeek V4.1 Flash | DeepSeek V4.1 Flash 0910 Uncensored: 2/5; DeepSeek V4.1 Flash Uncensored: 1/5 |
| MiMo V2.6 Flash | MiMo V2.6 Flash Uncensored: 2/5; MiMo V2.6 Flash Abliterated: 2/5 |
| GLM 5.3 | GLM 5.3 Uncensored: 2/5; Abliterated Large v2: 1/5 |
| Qwen3.5 27B | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted: 2/5; Qwen3.5 27B Queen Derestricted: 2/5; Qwen3.5 27B BlueStar v3 Derestricted: 2/5; Qwen3.5 27B Blossom V6.4 Derestricted: 2/5; Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted Lite: 1/5 |
So far the versions of one original land within a task of each other. With one run per task that gap is noise, so we cannot yet say whether abliteration, heretic edits or fine tunes cost tool use differently. More trials will tell.
Caveats
These are early numbers: one run per task, five tasks per model, so a single task is 20 points and small differences are noise. Tool calls are written in a text format rather than native function calling, which is harder on some models than others. We will add more trials and native tool calling, and update this post when we do. The live numbers are always on the rankings page.
Try any of them in a minute
No account and no sign up. Buy a prepaid key at openuncensored.com/credits, or hand your agent openuncensored.com/agents.md and let it get one for you. Then:
curl https://openuncensored.com/api/v1/chat/completions \
-H "Authorization: Bearer $OU_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "deepseek-v4.1-flash-1017-uncensored", "messages": [{"role": "user", "content": "Hello"}]}'
The full, always current table is on the rankings page.