Updated November 18, 2026
Uncensored LLM leaderboard.
34 uncensored and abliterated models you can call by API, ranked on how smart the original is, whether they finish agent tasks, whether they answer, how fast they run and how far they drifted.
Short answer
#1 DeepSeek V4.1 Flash 1017 Uncensored (69.8), #2 GLM 5.3 Flash Uncensored (69.7), #3 Qwen 3.8 27B Uncensored Fast by Heabsy (66.8). Scores are 0 to 100: intelligence 40%, agentic 30%, non-refusal 15%, speed 10%, low drift 5%. Every model is live on one OpenAI compatible API.The leaderboard
All 34 ranked uncensored models.
| # | Model | Score | Intelligence | Agentic | Answers | Speed | $ / 1M in · out |
|---|---|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 Uncensoredfrom DeepSeek V4.1 Flash 1017 | 69.8 | 67.5 | 55 (2/5) | 100% | 239 tok/s | $0.33 · $1.32 |
| 2 | GLM 5.3 Flash Uncensoredfrom GLM 5.3 Flash | 69.7 | 65.5 | 66 (3/5) | 94% | 178 tok/s | $0.12 · $0.50 |
| 3 | Qwen 3.8 27B Uncensored Fast by Heabsyfrom Qwen 3.8 27B | 66.8 | 62.2 | 49 (2/5) | 100% | 245 tok/s | $0.83 · $2.62 |
| 4 | DeepSeek V4.1 Flash 0910 Uncensoredfrom DeepSeek V4.1 Flash | 66.1 | 65.1 | 53 (2/5) | 97% | 209 tok/s | $0.33 · $1.32 |
| 5 | DeepSeek V4.1 Flash Uncensoredfrom DeepSeek V4.1 Flash | 64.3 | 65.1 | 40 (1/5) | 100% | 77 tok/s | $0.12 · $0.45 |
| 6 | Abliterated Large Flashfrom DeepSeek V4.1 Flash 1017 | 64.0 | 67.5 | 38 (1/5) | 92% | 184 tok/s | $1.00 · $2.00 |
| 7 | MiMo V2.6 Flash Uncensoredfrom MiMo V2.6 Flash | 63.5 | 62.3 | 55 (2/5) | 83% | 118 tok/s | $0.46 · $1.50 |
| 8 | Apex Flash 1 Abliteratedfrom GLM 5.3 Flash | 63.2 | 65.5 | 70 (3/5) | 86% | 10 tok/s | $0.15 · $0.50 |
| 9 | GLM 5.3 Uncensoredfrom GLM 5.3 | 63.0 | 67.0 | 53 (2/5) | 92% | 117 tok/s | $0.39 · $1.32 |
| 10 | MiMo V2.6 Flash Abliteratedfrom MiMo V2.6 Flash | 62.6 | 62.3 | 53 (2/5) | 83% | 99 tok/s | $0.50 · $1.50 |
| 11 | Abliterated Large v2from GLM 5.3 | 62.0 | 67.0 | 36 (1/5) | 94% | 283 tok/s | $3.00 · $5.00 |
| 12 | Qwen3.6 35B A3B Uncensored (HauhauCS Aggressive)from Qwen3.6 35B A3B | 62.0 | 53.7 | 49 (2/5) | 100% | 46 tok/s | $0.12 · $0.80 |
| 13 | Phala: Qwen3.8 27B Uncensored (Aggressive)from Qwen 3.8 27B | 62.0 | 62.2 | 49 (2/5) | 100% | 87 tok/s | $0.30 · $1.50 |
| 14 | Kimi K3.1 Pro Uncensoredfrom Kimi K3.1 Pro | 59.7 | 72.7 | 38 (1/5) | 97% | 59 tok/s | $2.50 · $5.00 |
| 15 | Gemma 4 31B Gembrain Uncensored Hereticfrom Gemma 4 31B | 58.5 | 54.7 | 66 (3/5) | 89% | 14 tok/s | $0.31 · $0.31 |
| 16 | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestrictedfrom Qwen3.5 27B | 57.8 | 54.8 | 51 (2/5) | 100% | 28 tok/s | $0.31 · $0.31 |
| 17 | Inkling 2 Small Uncensoredfrom Inkling 2 Small | 57.7 | 68.5 | 25 (0/5) | 94% | 395 tok/s | $0.60 · $2.40 |
| 18 | Qwen3.5 27B Queen Derestrictedfrom Qwen3.5 27B | 57.3 | 54.8 | 49 (2/5) | 100% | 28 tok/s | $0.31 · $0.31 |
| 19 | Qwen3.5 27B BlueStar v3 Derestrictedfrom Qwen3.5 27B | 56.4 | 54.8 | 49 (2/5) | 100% | 24 tok/s | $0.31 · $0.31 |
| 20 | Qwen3.5 27B Blossom V6.4 Derestrictedfrom Qwen3.5 27B | 56.3 | 54.8 | 49 (2/5) | 100% | 21 tok/s | $0.31 · $0.31 |
| 21 | Gemma 4 Uncensoredfrom Gemma 4 26B A4B non-reasoning | 54.6 | 41.5 | 53 (2/5) | 100% | 67 tok/s | $0.16 · $0.50 |
| 22 | Gemma 4 26B A4B Uncensored TEEfrom Gemma 4 26B A4B | 52.8 | 49.2 | 36 (1/5) | 100% | 21 tok/s | $0.15 · $0.70 |
| 23 | GLM 4.7 Flash Hereticfrom GLM 4.7 Flash | 52.7 | 32.9 | 51 (2/5) | 100% | 80 tok/s | $0.07 · $0.40 |
| 24 | GLM 4.6 Derestricted v5from GLM 4.6 | 50.2 | 46.2 | 38 (1/5) | 100% | 31 tok/s | $0.40 · $1.50 |
| 25 | Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted Litefrom Qwen3.5 27B | 49.8 | 54.8 | 34 (1/5) | 89% | 26 tok/s | $0.31 · $0.31 |
| 26 | Venice Uncensored 1.2from Mistral Small 24B Instruct 2501 | 49.3 | 25.0 | 49 (2/5) | 100% | 13 tok/s | $0.20 · $0.90 |
| 27 | Venice Role Play Uncensoredfrom Mistral Small 3.2 24B | 48.6 | 27.4 | 49 (2/5) | 92% | 51 tok/s | $0.50 · $2.00 |
| 28 | Qwen 2.5 32B Abliteratedfrom Qwen2.5 32B Instruct | 44.6 | 25.3 | 49 (2/5) | 100% | 19 tok/s | $0.70 · $0.70 |
| 29 | Llama 3.3 70B Instruct abliteratedfrom Llama 3.3 70B Instruct | 41.7 | 26.7 | 36 (1/5) | 100% | 19 tok/s | $0.70 · $0.70 |
| 30 | DeepSeek R1 Llama 70B Abliteratedfrom DeepSeek R1 Distill Llama 70B | 40.4 | 22.7 | 40 (1/5) | 100% | 18 tok/s | $0.70 · $0.70 |
| 31 | DeepSeek R1 Qwen Abliteratedfrom DeepSeek R1 Distill Qwen 32B | 40.0 | 33.0 | 23 (0/5) | 100% | 20 tok/s | $1.40 · $1.40 |
| 32 | Neural Daredevil 8B abliteratedfrom Llama 3 8B Instruct | 35.8 | 17.4 | 25 (0/5) | 100% | 32 tok/s | $0.44 · $0.44 |
| 33 | Llama 3 70B abliteratedfrom Llama 3 70B Instruct | 34.4 | 21.2 | 21 (0/5) | 100% | 17 tok/s | $0.70 · $0.70 |
| 34 | Aion-RP 1.0 (8B)from Llama 3.1 8B Instruct | 33.6 | 15.6 | 19 (0/5) | 100% | 36 tok/s | $0.80 · $1.60 |
By category
Leaders by category.
| Category | Model | Result |
|---|---|---|
| Best overallhighest OpenUncensored Score | DeepSeek V4.1 Flash 1017 Uncensored | score 69.8, rank #1 |
| Best for agents and codinghighest agentic score at 50+ tok/s | GLM 5.3 Flash Uncensored | agentic 66, 3/5 tool tasks finished, 178 tok/s |
| Fewest refusalshighest non-refusal in the top 10 | DeepSeek V4.1 Flash 1017 Uncensored | answered 100% of 36 edgy requests |
| Fastestmost tokens per second in the top 10 | Qwen 3.8 27B Uncensored Fast by Heabsy | 245 tok/s, rank #3 |
| Cheapest good modellowest price in the top 10 | DeepSeek V4.1 Flash Uncensored | $0.12 in, $0.45 out per 1M tokens, rank #5 |
| Best valuescore per dollar in the top 10 | GLM 5.3 Flash Uncensored | score 69.7 at $0.21 per 1M blended tokens |
| Smartest originalhighest intelligence (GPQA and HLE of the original) | Kimi K3.1 Pro Uncensored | intelligence 72.7, built from Kimi K3.1 Pro |
Method
How the score is built.
| Part | Weight | What it measures |
|---|---|---|
| Intelligence | 40% | Original model's published GPQA Diamond and HLE |
| Agentic | 30% | 5 tool scenarios: completion, no loops, no crashes |
| Non-refusal | 15% | 36 edgy requests answered, including finding people |
| Speed | 10% | Time for a 500 token answer: 2 s or less scores 100, 60 s or more 0 |
| Low drift | 5% | How close it stays to the original model |
Uncensoring removes refusals but can break a model: it loops in tool calls, crashes, or forgets what it knew. The score rewards models that stay smart and reliable while answering. Details, prompts and limits are on the methodology page.
Not ranked yet
Scored, but not ranked.
The original model has no published GPQA Diamond or HLE result, so the intelligence part is missing and the score is partial.
| Model | Partial score | Agentic | Answers | Speed |
|---|---|---|---|---|
| Aion 3.5 Mini | 78.6 | 83 | 94% | 58 tok/s |
| DeepSeek V4.1 Flash Heretic | 72.7 | – | 94% | 53 tok/s |
| Abliterated Model | 68.4 | 49 | 100% | 114 tok/s |
| Aion 3.5 | 63.6 | 55 | 92% | 51 tok/s |
| Gemma 4 26B A4B Chimera X Heretic | 61.9 | – | 100% | 43 tok/s |
| Gemma 4 31B SDFT Heretic RP | 61.3 | – | 100% | 10 tok/s |
| CyberHeabsy | 59.9 | 38 | 94% | 146 tok/s |
| Gemma 4 31B Cognitive Unshackled | 59.6 | – | 100% | 19 tok/s |
| Aion 3.0 Mini | 56.1 | 38 | 92% | 59 tok/s |
| Aion 3.0 | 44.3 | 21 | 83% | 26 tok/s |
Use the data
Download the leaderboard.
- JSON with live speed, free and without a key:
https://openuncensored.com/api/rankings.json - Markdown for agents and LLMs: /uncensored-llm-leaderboard.md
- Over MCP: the
get_rankingstool of the OpenUncensored MCP server.
Choosing an API rather than a model? See the best uncensored LLM API and the cheapest uncensored LLM API.
FAQ
Questions people ask.
What is the best uncensored LLM right now?
DeepSeek V4.1 Flash 1017 Uncensored leads the OpenUncensored leaderboard with 69.8, ahead of GLM 5.3 Flash Uncensored (69.7) and Qwen 3.8 27B Uncensored Fast by Heabsy (66.8).
How is the uncensored LLM leaderboard scored?
Each model gets a 0 to 100 score: intelligence 40%, agentic 30%, non-refusal 15%, speed 10%, low drift 5%. Non-refusal is measured on 36 requests censored models refuse, agentic on 5 tool scenarios, speed on real API traffic or our own speed test.
Which uncensored model refuses the least?
21 of the 34 ranked models answered all 36 test requests. The highest ranked of them is DeepSeek V4.1 Flash 1017 Uncensored.
Can I download the leaderboard data?
Yes, free and without a key: https://openuncensored.com/api/rankings.json (JSON, with live speed) or https://openuncensored.com/uncensored-llm-leaderboard.md (markdown).
How is this different from other uncensored leaderboards?
Most uncensored leaderboards measure whether a model answers. This one ranks models you can call by API today and also measures whether they finish agent tasks, how smart the original model is, how fast they run on real traffic, and what they cost.
Updated November 18, 2026. Scores measured November 8, 2026.