Arli AI sells “unrestricted inference” at a flat monthly rate, with no per-token billing. It runs everything on its own servers and makes its own Derestricted models with a technique it calls norm-preserving biprojected abliteration. We list 131 uncensored models from it: its two Derestricted bases plus a long tail of roleplay and writing fine-tunes. Only NanoGPT has more in our index.
This review is based on Arli AI’s published pricing, model list, and policies, checked on October 11, 2026. Benchmark results from our own runs will be added when they’re complete. Live numbers are on the Arli AI provider page.
The short version
| Best for | Heavy roleplay and writing use at a fixed monthly cost |
| Uncensored models | 131 (103 fine-tuned, 2 abliterated, 26 undisclosed) |
| Per-token price | None: flat-rate plans only |
| Plans | Free (limited), Starter $10/mo, Core $15, Plus $20, Pro $30, Max $160 |
| Privacy score | 33 / 100, joint 22nd of 37 |
| Watch out for | Your plan caps context length, and payment methods aren’t named publicly |
Models
Arli’s own work is two abliterated models:
- GLM-4.6-Derestricted-v5: a 355B abliteration of GLM-4.6, with a 128K context window. It’s the only listed model that needs the Core plan or higher.
- Qwen3.5-27B-Derestricted: a 27B abliteration of Qwen3.5, with a 256K context window.
Both have spread to other hosts. NanoGPT resells GLM 4.6 Derestricted v5 at $0.40 / $1.50 per 1M tokens, and Featherless serves Qwen3.5-27B-Derestricted at $1.60 / $12.
The rest of the catalog is fine-tunes:
- 44 models built on Qwen3.5-27B-Derestricted. These inherit Arli’s abliteration, then add a roleplay or writing tune.
- 61 Gemma 4 31B tunes, such as Artemis, Gembrain, MeroMero, Novelist and a few Heretic variants.
- 24 Qwen 3.8 27B tunes, such as DarkIdol, Kimiko, Synthia and Yes-Man-uncensored.
That’s a lot of variety at two model sizes. If you want something much bigger than 31B, the only uncensored choice here is the 355B GLM. Arli also serves models we don’t list as uncensored, such as GLM-4.7 and DeepSeek-V4-Flash.
Most of the Qwen 3.8 tunes don’t say how they were uncensored, so 26 models are marked “Unverified” until we test them. Search them all on the Models page.
Pricing
All paid plans are flat-rate with unlimited tokens and requests. You pay for model size, context length and parallel requests:
| Plan | Price | Models | Context | Parallel requests |
|---|---|---|---|---|
| Free | $0 | Limited | 12K | 1, delayed |
| Starter | $10/mo | Up to 31B | 16K | 1 |
| Core | $15/mo | Up to 355B | 32K | 1 |
| Plus | $20/mo | Up to 355B | 128K | 1 |
| Pro | $30/mo | Up to 355B | 256K | 2 |
| Max | $160/mo | Up to 355B | 512K | 6 |
Starter covers every uncensored model we list except GLM-4.6-Derestricted-v5. Annual billing is 10% off. Arli’s docs say free accounts are limited to 5 requests per model every 2 days, so the free tier is for trying a model, not for using it.
The context column matters. Many models support 256K, but on Starter you get 16K: enough for chat, short for long roleplay sessions or documents. Plus ($20) is the cheapest plan with 128K.
Because there’s no per-token price, we can’t give a cost per 1,000 chats. For a rough comparison: NanoGPT sells many of the same Gemma 4 tunes at $0.306 per 1M tokens, so $10 there buys about 33M tokens. Starter is the better deal if you use more than that each month.
Privacy
Arli AI scores 33 / 100 on our privacy rubric, which only gives points for documented claims:
| Rubric item | Arli AI | Points |
|---|---|---|
| Prompt retention | Doesn’t store, retain, or have access to prompts or outputs | 20 / 20 |
| No training on prompts | Not stated | 0 / 15 |
| Technical guarantee (TEE / E2EE) | None; relies on policy | 0 / 20 |
| Signup | Account required | 3 / 8 |
| Crypto payments | Not documented | 0 / 7 |
| Jurisdiction | Not stated | 0 / 10 |
| Upstream hosts disclosed | Yes: all inference runs on its own servers | 10 / 10 |
| Independent audit | No | 0 / 10 |
The strongest point is that there’s no hidden upstream. Arli says all inference runs in-house, so its zero-log statement covers the whole path. Many resellers can’t say that. It keeps usage metadata for up to 60 days.
The points it loses are mostly gaps in documentation: its privacy policy doesn’t address training use or name a jurisdiction, and it doesn’t say which payment methods it takes.
Access and payment
- Web chat and an OpenAI-compatible API (chat and text completions).
- Account required.
- Payment methods aren’t named publicly. The terms only mention “third-party payment processors”.
- The terms also offer one-time (non-recurring) payment, prorated refunds and a 30-day satisfaction guarantee.
How it compares
- Other flat-rate hosts: Infermatic starts at $9/mo for roleplay fine-tunes up to 72B (privacy score 40), but doesn’t claim abliterated weights. Featherless has a $25/mo unlimited chat plan with a much larger Hugging Face catalog, capped at 32K context.
- Pay per token instead: NanoGPT resells many of the same Gemma 4 tunes at $0.306 / $0.306 per 1M tokens, with no account required.
- More private on paper: Venice (55) states it doesn’t train on prompts and names its jurisdiction.
Verdict
Arli AI is a good fit if you use uncensored models a lot and want a fixed bill. It makes its own abliterated bases, runs them on its own hardware, and gives you more than a hundred tunes built on top of them. The $10 Starter plan covers almost all of it.
Check the context limit before you pick a plan: the cheap tiers are tight. And if documented training and payment policies matter to you, Arli’s published pages don’t cover them yet.
Compare every Arli AI model on its provider page, or see every provider in the index.