Uncensored models for every job — 7B to 70B.

The UM family is a set of uncensored language models that share one API, one toolset and one behaviour: answer the question. Pick by speed, context and depth.

Join the waitlist Use the API

Lineup

All uncensored models at a glance

Model	Best for	Context	Speed	Status
UM-7B uncensored-small	Edge, chat, high-volume classification	32K	~180 tok/s	Live
UM-14B uncensored-medium	Balanced assistants, tool-use agents	64K	~120 tok/s	Live
UM-32B uncensored-pro	Research, long-context reasoning, RAG	128K	~74 tok/s	Live
UM-70B uncensored-flagship	Creative, companion & roleplay depth (18+)	128K	~46 tok/s	Live
UM-32B-R reasoning-pro	Transparent chain-of-thought agents	128K	~60 tok/s	Live
UM-70B-R reasoning-flagship	Flagship reasoning with full traces	128K	~40 tok/s	Preview

Choosing

Which uncensored model should you use?

UM-7B · fastest

~180 tok/s and 32K context. Great for high-volume chat, classification, moderation pipelines and anything latency-sensitive. Included free.

UM-14B · best value

The sweet spot for assistants and tool-using agents. Noticeably smarter than 7B, still fast, still on the free tier.

UM-32B · all-rounder

128K context for long documents, RAG and research. The default for production apps on Pro.

UM-70B · flagship

The richest writing and deepest knowledge — built for fiction, companions and roleplay that needs to feel alive.

UM-32B-R · reasoning

Shows its chain-of-thought before answering. Use it for math, planning, code and agents you want to audit.

UM-70B-R · preview

Flagship reasoning with full traces. Currently in preview for selected accounts.

How they're made

From safety-tuned base to uncensored model

Distil. We build a high-signal dataset of thorough, accurate answers to sensitive-but-legitimate prompts — the ones mainstream models refuse.
Fine-tune. Open base models are trained with SFT + DPO to prefer the helpful answer over the refusal, without lobotomising general ability.
Reason. The -R variants learn transparent reasoning via GRPO, so you can read how they reached an answer.
Evaluate. Every release is tested for both refusal rate and quality retention against its base — uncensored is worthless if it gets dumber.

Want the background? Read how uncensored LLMs work, or compare with open-weight options in best uncensored LLMs.

FAQ

Frequently asked questions

What makes a model “uncensored”?

It has been trained to stop refusing legitimate requests that safety-tuned models decline. It still has knowledge, reasoning and style from its base model — just without the reflexive refusals. Details in What is an uncensored LLM?

Which uncensored model is best?

For most chat and agent use, UM-14B is the best value. UM-32B is the best all-rounder for long documents and research; UM-70B for creative depth and roleplay; UM-32B-R when you need step-by-step reasoning. Looking at open-weight options too? See best uncensored LLMs.

How were the UM models trained?

We distil high-quality responses into a curated dataset, fine-tune open base models with SFT + DPO toward helpfulness on sensitive-but-legal topics, and add transparent reasoning via GRPO for the -R models.

Can I download the weights?

The hosted models are API-only. Enterprise customers can get private fine-tunes and dedicated deployments.

What does “Coverage” or refusal rate mean?

We measure how many prompts from a benchmark of legitimate-but-sensitive requests a model answers helpfully instead of refusing. Larger UM models answer roughly 97–99% of them.

Coming soon

Get early access to uncensored AI that actually answers.

We're opening in waves. Join the waitlist and we'll email you when your spot is ready — nothing else.