Uncensored LLMs on one API key.
14 loosely aligned and refusal-stripped models through a single OpenAI-compatible endpoint, priced per token, with a hard spend ceiling on every key and no prompt retention.
What uncensored means here, precisely
The word gets used for two different things, and the difference decides whether a model will actually finish your answer.
An uncensored model is a finetune whose alignment was loosened during training. It refuses far less than a frontier model, and it can still refuse. An abliterated model had the refusal direction removed from its weights outright, so it very nearly cannot. MagmaRouter tags the two separately rather than blurring them, because a page that calls everything uncensored is not telling you which one you are buying.
The catalogue
| Model | Context | In | Out | Tools |
|---|---|---|---|---|
| Magnum v4 72B anthracite-org/magnum-v4-72b |
33k | $2.88 | $5.75 | No |
| Venice: Uncensored cognitivecomputations/dolphin-mistral-24b-venice-edition |
128k | $0.230 | $1.04 | No |
| MythoMax 13B gryphe/mythomax-l2-13b |
8k | $0.092 | $0.127 | No |
| Mancer: Weaver (alpha) mancer/weaver |
8k | $0.460 | $0.863 | No |
| Nous: Hermes 3 405B Instruct nousresearch/hermes-3-llama-3.1-405b |
131k | $1.15 | $1.15 | No |
| Nous: Hermes 3 70B Instruct nousresearch/hermes-3-llama-3.1-70b |
131k | $0.805 | $0.805 | No |
| Nous: Hermes 4 405B nousresearch/hermes-4-405b |
131k | $1.15 | $3.45 | No |
| Sao10K: Llama 3 8B Lunaris sao10k/l3-lunaris-8b |
8k | $0.046 | $0.057 | No |
| Sao10K: Llama 3.1 Euryale 70B v2.2 sao10k/l3.1-euryale-70b |
131k | $0.977 | $0.977 | Yes |
| Sao10K: Llama 3.3 Euryale 70B sao10k/l3.3-euryale-70b |
131k | $0.748 | $0.863 | No |
| TheDrummer: Cydonia 24B V4.1 thedrummer/cydonia-24b-v4.1 |
131k | $0.345 | $0.575 | No |
| TheDrummer: Skyfall 36B V2 thedrummer/skyfall-36b-v2 |
33k | $0.632 | $0.920 | No |
| TheDrummer: UnslopNemo 12B thedrummer/unslopnemo-12b |
1M | $0.460 | $0.460 | No |
| ReMM SLERP 13B undi95/remm-slerp-l2-13b |
6k | $0.403 | $0.748 | No |
One base URL
Anything speaking the OpenAI chat completions format works without a code change.
curl https://api.magmarouter.com/v1/chat/completions \
-H "Authorization: Bearer rl-..." \
-H "Content-Type: application/json" \
-d '{
"model": "sao10k/l3-lunaris-8b",
"messages": [{"role": "user", "content": "Hello"}]
}'
FAQ
What counts as an uncensored model?
A finetune whose alignment was loosened by training, so it answers a great deal that a frontier model declines. It is not a guarantee: an uncensored model can still refuse, because the refusal behaviour was reduced rather than removed. The stronger claim is abliteration.
Do I need my own GPU?
No. These are the same open weights you could run locally, served per token. A 70B build at usable quality wants roughly 48GB of VRAM to self-host, which is a card you buy or a GPU you rent by the hour whether or not you use it.
Will it write anything at all?
No, and anyone promising that is selling something. A loosely aligned model answers far more than a frontier one, and it is also more confidently wrong. Removing a refusal does not add knowledge, and it does not change what you are allowed to do with the answer.
Can I cap what a key spends?
Yes. The cap is enforced before the request leaves, so the reply cannot cost more than the hold taken before it was sent. That is the difference between a budget and a limit.