Best Uncensored LLMsWhich uncensored LLMs actually finish agent tasks?
Uncensoring can break tool use. We ran every ranked uncensored model through five multi-step agent tasks: how many finish, how many loop, and which ones to use in Cline, Codex or OpenCode.
An uncensored model that refuses less is no use in a coding agent if it loses track of its tools. Removing refusals changes the model's weights, and some edits cost more tool-use reliability than others. Every ranked model gets the same five multi-step tool tasks, one run each, scored on finishing without looping or crashing.
Short answerNo ranked model finishes more than 3 of the 5 tasks yet. 19 loop on 40% or more of them. Start with Apex Flash 1 Abliterated, then GLM 5.3 Flash Uncensored.
Best for agents
| Rank | Model | Score | Agent tasks | Loops | Answers edgy | Hard-line refusals | $ / 1M in+out | Where to run it |
|---|---|---|---|---|---|---|---|---|
| 8 | Apex Flash 1 Abliterated | 63.2 | 3/5 | 0% | 86% | 100% | $0.65 | NanoGPTOpenUncensored |
| 2 | GLM 5.3 Flash Uncensored | 69.7 | 3/5 | 40% | 94% | 14% | $1.00 | NanoGPTOpenUncensored |
| 14 | Gemma 4 31B Gembrain Uncensored Heretic | 58.5 | 3/5 | 40% | 89% | 0% | $0.61 | NanoGPTOpenUncensored |
| 1 | DeepSeek V4.1 Flash 1017 Uncensored | 69.8 | 2/5 | 0% | 100% | 0% | $1.65 | VeniceNanoGPTOpenUncensored |
| 7 | MiMo V2.6 Flash Uncensored | 63.5 | 2/5 | 0% | 83% | 93% | $2.00 | NanoGPTOpenUncensored |
| 4 | DeepSeek V4.1 Flash 0910 Uncensored | 66.1 | 2/5 | 20% | 97% | 0% | $1.65 | NanoGPTOpenUncensored |
| 9 | GLM 5.3 Uncensored | 63.0 | 2/5 | 20% | 92% | 0% | $3.50 | NanoGPTOpenUncensored |
| 10 | MiMo V2.6 Flash Abliterated | 62.6 | 2/5 | 20% | 83% | 93% | $2.00 | NanoGPTOpenUncensored |
| 19 | Gemma 4 Uncensored | 54.6 | 2/5 | 20% | 100% | 0% | $0.66 | VeniceOpenUncensored |
| 15 | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | 57.8 | 2/5 | 40% | 100% | 0% | $0.61 | NanoGPTOpenUncensored |
What we saw
- Average tasks finished across all ranked models: 1.6 of 5.
- Average loop rate: 35%. Loops are the usual failure: the model repeats a tool call instead of moving on.
- One run per task means a single task of difference is within noise. Treat close scores as ties.