Best abliterated models and APIs, ranked
The best abliterated model is Kimi K3.1 Pro Abliterated, and the best abliterated API is Silk Compute at 89.2 out of 100. To run one at home, Qwen 3.8 27B Abliterated fits a single 24 GB GPU.
Top picks
Latest reports
Nov 18, 2026
The November 2026 Abliterated API Index
| Measure | Value | Rank | |
|---|---|---|---|
| Abliterated API Index | 89.2 | 1/6 | |
| Price | $0.405 / 1M | 1/6 | |
| Prompt retention | Never stored | 1/6 | |
| Context window | 1M tokens | 1/6 | |
| 1M-context models | 5 models | 1/6 | |
| API formats | 10 / 10 | 1/6 | |
| Payment and access | 7 / 10 | 2/6 |
Best abliterated models
| # | Model | Parameters | Context | Runs on | Lowest price per 1M (in / out) | Best for |
|---|---|---|---|---|---|---|
| 1 | Kimi K3.1 Pro AbliteratedMoonshot AI | 2,780B | 1M | Hosted API | $3.00 / $8.00 | Best overallSecurity research |
| 2 | DeepSeek V4.1 Flash 1017 AbliteratedDeepSeek | 763B | 1M | Hosted API | $0.24 / $0.90 | Best valueCoding agents |
| 3 | GLM-5.3 AbliteratedZ.ai | 753B | 1M | Hosted API | $2.00 / $4.00 | CodingLong documents |
| 4 | GLM-5.3 Flash AbliteratedZ.ai | 320B · 18B active | 1M | Hosted API | $0.24 / $1.00 | Lowest input priceHigh volume |
| 5 | Qwen 3.8 27B AbliteratedAlibaba, abliterated by huihui-ai | 27.8B | Varies | 24 GB GPU | $1.60 / $12.00 | Best local model24 GB GPU |
| 6 | Gemma 4 31B HereticGoogle, abliterated with Heretic | 31.3B | Varies | 24 GB GPU | Free weights | 24 GB GPUWriting |
| 7 | Qwen 3.6 35B-A3B AbliteratedAlibaba, abliterated by huihui-ai | 36B · 3B active | Varies | 24 GB GPU | Free weights | Fast on a home GPUMixture of experts |
| 8 | Llama 3.1 8B AbliteratedMeta, abliterated by mlabonne | 8B | Varies | 8 GB GPU or laptop | Free weights | Laptop8 GB GPU |
Rankings by job
Best abliterated model for coding agents
Kimi K3.1 Pro Abliterated is the best abliterated model for coding agents, and DeepSeek V4.1 Flash Abliterated is the best value.
See the full rankingEvery abliterated API, side by side
Read from each provider's pages on November 18, 2026.
| # | Provider | Best-value 128K+ model | Price per 1M (in / out) | Context | Prompts stored | Anthropic API | Crypto | Index |
|---|---|---|---|---|---|---|---|---|
| 1 | Silk Compute | DeepSeek V4.1 Flash 1017 | $0.24 / $0.90 | 1M | Never stored | Yes | Yes | 89.2 |
| 2 | Abliteration | Abliterated Large Flash | $1.00 / $2.00 | 1M | Never stored | Yes | Not listed | 68.2 |
| 3 | Morph | DeepSeek V4.1 Flash Uncensored | $0.30 / $1.20 | 1M | Up to 30 days | Yes | USDC | 57.1 |
| 4 | Unfil.AI | Hermes 3 405B | $1.20 / $1.20 | 128K | Deleted within 24 h | Not stated | Yes | 35.4 |
| 5 | Awan LLM | None | $5 to $80 a month | 8K | Never stored | Not stated | PayPal only | 30.1 |
| 6 | Featherless | None | None | 32K | Never stored | Not stated | No | 27.5 |
How the index works
Six measures, with the same weights for every provider.
Price
Prompt retention
Context window
1M-context models
API formats
Payment and access
Questions
Is an abliterated model the same as an uncensored model?
Abliteration is one way to make an uncensored model. It removes the refusal direction without retraining, where an uncensored fine-tune retrains the model.
Do abliterated models lose quality?
A little, on some builds. Large models lose less, which is one reason the 700B-class hosted builds rank highest.
Do abliterated models still refuse?
Sometimes. Some builds still decline in a few categories.
What are abliterated models used for?
Coding agents, security research such as malware analysis, authorized red teaming, and training data for safety classifiers.