Best Uncensored LLMsEvery uncensored DeepSeek V4.1 Flash, compared: 1017, 0910, Heretic and thinking
There are at least eight uncensored DeepSeek V4.1 Flash versions by API. Which release each is built on, how they score on refusals and agent tasks, context, price at the cheapest host, and which to use.
DeepSeek V4.1 Flash is the most used uncensored model family on OpenUncensored, and there are now several uncensored versions of it, built on different DeepSeek releases with different methods. Here is every version you can call by API.
Ranked versions
| Rank | Model | Score | Agent tasks | Loops | Answers edgy | Hard-line refusals | $ / 1M in+out | Where to run it |
|---|---|---|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 Uncensored | 69.8 | 2/5 | 0% | 100% | 0% | $1.65 | VeniceheabsyNanoGPTOpenUncensored |
| 4 | DeepSeek V4.1 Flash 0910 Uncensored | 66.1 | 2/5 | 20% | 97% | 0% | $1.65 | NanoGPTOpenUncensored |
| 5 | DeepSeek V4.1 Flash Uncensored | 64.3 | 1/5 | 0% | 100% | 0% | $0.57 | VenicerefuselessNanoGPTOpenUncensored |
Every version by API
| Model id | Built on | Context | $ / 1M in · out | Hosts |
|---|---|---|---|---|
| deepseek-v4.1-flash-0910-uncensored | – | 524K | $0.33 · $1.32 | NanoGPT |
| deepseek-v4.1-flash-0910-uncensored-thinking | – | 524K | $0.33 · $1.32 | NanoGPT |
| deepseek-v4.1-flash-1017-heretic | – | 512K | $0.20 · $0.79 | NanoGPT |
| deepseek-v4.1-flash-1017-heretic-thinking | – | 512K | $0.20 · $0.79 | NanoGPT |
| deepseek-v4.1-flash-1017-uncensored | DeepSeek-V4.1-Flash-1017 | 1M | $0.33 · $1.32 | Venice, heabsy, NanoGPT |
| deepseek-v4.1-flash-1017-uncensored-thinking | – | 1M | $0.33 · $1.32 | NanoGPT, Venice |
| deepseek-v4.1-flash-uncensored | DeepSeek-V4.1-Flash | 1M | $0.12 · $0.45 | Venice, Refuseless, NanoGPT |
| deepseek-v4.1-flash-uncensored-thinking | – | 1M | $0.33 · $1.32 | NanoGPT, Venice |
How the versions differ
- 1017 vs 0910: DeepSeek's October 17 and September 10 releases of V4.1 Flash. The 1017 Uncensored build ranks above the 0910 one and has the full 1M context.
- Heretic: made with Heretic, which automates abliteration and searches for the smallest edit that removes refusals. The 1017 Heretic build has a high score but no agent-task results yet, and a shorter context.
- Thinking: the same model with reasoning turned on. Better on hard problems, more output tokens.
Open weights
If you want to run it yourself, OpenUncensored published its own uncensored 1017 weights on Hugging Face: MMLU 87.89% against 88.16% for the original, and 37 of 48 test requests fully answered against 9. They are not the build the APIs above serve, and they need a large multi-GPU server (510 GB).
Frequently asked questions
Which uncensored DeepSeek V4.1 Flash is best?
As of November 2026, the 1017 Uncensored build is the top ranked uncensored model overall in these rankings, answers every edgy test request and has a 1M context. The 1017 Heretic build scores higher but has not been through the agent tasks yet.
What do 1017 and 0910 mean?
They are DeepSeek's dated V4.1 Flash releases: 1017 is the newer one. Uncensored builds are made from each, so check which release a build is based on.
What is the thinking version?
The same model with reasoning turned on: it thinks before it answers, which helps on harder problems and costs more output tokens.