Featherless.ai is a serverless host for open-weight models from Hugging Face, behind an OpenAI-compatible API. Its public model API listed 22,182 models when we checked. About 1,000 of them matched uncensored, abliterated, Heretic, Dolphin, Hermes or Derestricted, and 277 had “abliterated” in the name. That’s the largest catalog of community uncensored checkpoints we found. We list 25 popular picks by downloads; you can search the rest on Featherless itself.
This review is based on Featherless’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 Featherless provider page.
The short version
| Best for | Running a specific community abliterated or Hermes checkpoint without hosting it yourself |
| Uncensored models | 25 listed (18 abliterated, 7 fine-tuned), out of about 1,000 on the platform |
| Per-token price | $0.05 / $0.08 to $1.60 / $12 per 1M tokens (input / output) |
| Plans | Chat $25/mo (unlimited tokens), Developer from $50/mo (credit, per token) |
| Privacy score | 23 / 100, 28th of 35 |
| Watch out for | 32K context on every model we list, and a short privacy policy |
Models
Featherless serves Hugging Face repos under their original names, so you know exactly which checkpoint you’re getting. Our 25 picks cover the main families:
- Abliterated Gemma 4 and Qwen. huihui-ai’s Gemma 4 31B abliterated v2 ($0.12 / $0.36), a Gemma 4 31B Heretic, a Qwen3.6 35B-A3B abliteration ($0.40 / $4), and Qwen3.8 27B abliterated and Heretic builds.
- Arli AI’s Qwen3.5-27B-Derestricted ($1.60 / $12), the same base Arli sells on flat-rate plans.
- Llama abliterations. mlabonne’s Llama 3.1 8B abliterated ($0.05 / $0.08) and Llama 3.1 70B lorablated ($2.88 / $2.88), and huihui-ai’s Llama 3.3 70B abliterated ($0.65 / $0.75).
- Hermes and Dolphin fine-tunes. Hermes 3 (8B and 70B), Hermes 4 (14B and 70B), Dolphin 3.0 Llama 3.1 8B, and Dolphin-Mistral-24B-Venice-Edition ($0.70 / $1.16), the same weights as Venice’s own model.
All 25 have a documented method: 18 are abliterated and 7 are fine-tuned. Search them on the Models page.
The limit is context. Every model we list is served at 32K through the API, even where the base model supports far more. Featherless’s pricing page says the Developer plan goes “up to 256K”, but our model data shows 32K across the board.
Pricing
There are two ways to pay.
The Chat plan is $25/mo with unlimited tokens, 32K context and 4 concurrency units. Each model uses 1 to 4 units depending on size, so a big model can take your whole allowance.
The Developer plan starts at $50/mo of credit, billed per token at each model’s rate. Only successful requests are billed, and credits don’t expire. A typical short chat (1,000 tokens in, 500 out) costs:
- about $0.09 per 1,000 chats on Llama 3.1 8B abliterated
- about $0.30 per 1,000 chats on Gemma 4 31B abliterated v2
- about $1.03 per 1,000 chats on Llama 3.3 70B abliterated
- about $7.60 per 1,000 chats on Qwen3.5-27B-Derestricted or the Qwen3.8 27B abliterations ($1.60 / $12)
Prices don’t track model size closely. Some 27B Qwen models cost more than the 70B Llamas, and Dolphin-Mistral-24B-Venice-Edition costs about twice Venice’s own $0.20 / $0.90. Check the per-model rate before you commit.
There’s no free tier on the pricing page.
Privacy
Featherless scores 23 / 100 on our privacy rubric, which only gives points for documented claims:
| Rubric item | Featherless | Points |
|---|---|---|
| Prompt retention | Doesn’t collect, store, or log API prompts or completions | 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 | No (card via Stripe) | 0 / 7 |
| Jurisdiction | Not stated | 0 / 10 |
| Upstream hosts disclosed | No | 0 / 10 |
| Independent audit | No | 0 / 10 |
The no-logging statement is clear: its privacy page says API requests are processed in real time and not stored. Most of the missing points come from what the page doesn’t say. It doesn’t address training use or name a jurisdiction, and the pages we checked don’t name a legal entity.
Featherless also matters beyond its own score. Hugging Face Inference Providers routes almost all of its uncensored models to Featherless, so this policy is the one that applies there too.
Access and payment
- OpenAI-compatible API, plus a chat interface.
- Account required.
- Card payments via Stripe. Business plans can be invoiced.
- The company grew out of the RWKV open-source model project.
How it compares
- Cheaper small models: DeepInfra and OpenRouter serve 8B roleplay models from $0.04 / $0.05 per 1M tokens, with better privacy scores (58 and 65). Neither lists abliterated checkpoints.
- Same checkpoints, different terms: Hugging Face (58) reaches the same Featherless-served models through its own router, with no markup on provider rates.
- Longer context: NanoGPT serves several abliterated and Derestricted models with 256K context or more, no account needed.
Verdict
Featherless is the place to go when you want a specific Hugging Face checkpoint by name and don’t want to run a GPU. Nobody else in our index comes close on the long tail of abliterated and Heretic models, and the cheapest 8B abliteration costs almost nothing.
The trade-offs are the 32K context, uneven per-model pricing, and a privacy policy that covers logging but not much else. If you only need one popular model, a host that sells it directly may be cheaper.
Compare our Featherless picks and prices on its provider page, or see every provider in the index.