What DeepSeek abliterated models exist?
Three groups of models get called “DeepSeek abliterated”: abliterated R1 distills you can run locally, abliterated or uncensored builds of newer DeepSeek models that providers host, and a few community edits of the full-size models. The first group is by far the most downloaded.
DeepSeek released R1 in January 2025 together with six smaller “distilled” models. The R1 paper and model card describe the full model as a mixture-of-experts network with 671B total parameters, 37B of them active for each token, and a 128K context window. It was trained from DeepSeek-V3-Base, and the weights are MIT licensed.
Soon after the release, the Hugging Face account huihui-ai published abliterated versions of the distills, including DeepSeek-R1-Distill-Qwen-32B-abliterated and DeepSeek-R1-Distill-Llama-8B-abliterated. They are also packaged for Ollama as huihui_ai/deepseek-r1-abliterated, with tags from 1.5b to 70b. That page lists hundreds of thousands of pulls. Later releases followed the same path: an abliterated DeepSeek-R1-0528-Qwen3-8B appeared after DeepSeek’s May 2025 update to R1.
On the hosted side, this directory lists 4 DeepSeek-based models across Refuseless, Uncensored.com and UncensoredLLM. One of them is labeled abliterated: DeepSeek V4.1 Flash 1017 Abliterated on Refuseless. The rest are uncensored builds or plain model names. They are covered in a table further down.
If abliteration itself is new to you, read what abliterated models are first. This post assumes you know the basic idea: a “refusal direction” is found in the model’s activations and removed from its weights.
Is DeepSeek-R1-Distill the same as DeepSeek R1?
No. A DeepSeek-R1-Distill model is a different, much smaller model that learned to imitate R1’s reasoning style. DeepSeek generated reasoning samples with R1 and used them to fine-tune existing open models. The architecture, knowledge and license come from the base model, not from R1.
| Distilled model | Base model | License to check |
|---|---|---|
| DeepSeek-R1-Distill-Qwen-1.5B | Qwen2.5-Math-1.5B | Apache 2.0 (Qwen) |
| DeepSeek-R1-Distill-Qwen-7B | Qwen2.5-Math-7B | Apache 2.0 (Qwen) |
| DeepSeek-R1-Distill-Llama-8B | Llama-3.1-8B | Llama 3.1 license |
| DeepSeek-R1-Distill-Qwen-14B | Qwen2.5-14B | Apache 2.0 (Qwen) |
| DeepSeek-R1-Distill-Qwen-32B | Qwen2.5-32B | Apache 2.0 (Qwen) |
| DeepSeek-R1-Distill-Llama-70B | Llama-3.3-70B-Instruct | Llama 3.3 license |
This matters for three reasons. First, quality: a 7B or 8B distill reasons in R1’s style but has a fraction of its knowledge, and it shows on hard or niche prompts. Second, licensing: the Llama-based distills carry Meta’s Llama license on top of DeepSeek’s MIT terms, so check it before you ship weights or build a product on them. Third, naming: the Ollama tag deepseek-r1-abliterated:8b is a Llama model under the hood, and the 14b and 32b tags are Qwen models.
An abliterated distill is two steps away from R1: distilled first, then edited. When a listing just says “DeepSeek R1 abliterated”, find out which of these you’re getting.
How does abliteration work on a reasoning model?
The method is the same as for any chat model: collect activations on prompts the model refuses and on harmless ones, find the direction that separates them, and remove it from the weights. The difference is where refusals show up. A reasoning model can decide to decline inside its thinking, before it writes a single visible word.
That creates a few behaviors you won’t see with ordinary chat models:
- Refusals inside the reasoning. The thinking can drift toward policy talk (“I should be careful here”) and end in a soft refusal or a watered-down answer, even when the final answer looks cooperative.
- Skipped thinking. The huihui-ai model card for the abliterated Qwen-32B distill notes that the model sometimes doesn’t produce a
<think>tag or still refuses, and suggests giving it an example first and then asking the question. - Weaker reasoning on hard tasks. Abliteration edits every layer, and multi-step reasoning is one of the first skills to suffer from a blunt edit. Check math, logic and long instructions, not just whether it refuses.
One more thing to know: DeepSeek’s models also avoid some politically sensitive topics, and that behavior isn’t always the same thing as a general refusal. Perplexity released a post-trained R1 called R1 1776 aimed at that specific issue, using fine-tuning rather than abliteration. If your concern is adult fiction or roleplay rather than politics, a general abliteration or an uncensored fine-tune is the relevant fix.
How much do thinking tokens cost on a DeepSeek model?
Thinking tokens are usually billed at the output rate, and output costs more than input. On a reasoning model, the price list understates the bill, because you pay for reasoning text your users never see.
How much reasoning a model writes depends on the prompt. DeepSeek’s own model card for R1-0528 says the updated model averaged about 23,000 tokens per question on the AIME 2025 math test, up from about 12,000 for the original R1. Chat and roleplay prompts need far less than competition math, but even a few hundred to a few thousand thinking tokens per reply adds up.
Here is the effect on a roleplay turn of 6,000 input tokens and 300 visible output tokens, at three thinking budgets. The budgets are illustrative, not measurements: log your own to replace them.
| Provider and model | No thinking | +1,000 thinking | +4,000 thinking |
|---|---|---|---|
| RefuselessDeepSeek V4.1 Flash 1017 Abliterated · $0.12 / $0.45 | $0.855 | $1.305 | $2.655 |
| Uncensored.comDeepSeek R1 0528 · $0.50 / $2.18 | $3.654 | $5.834 | $12.374 |
| Uncensored.comDeepSeek V4.1 Flash Cyber Uncensored · $0.96 / $1.92 | $6.336 | $8.256 | $14.016 |
| UncensoredLLMDeepSeek V4 Pro Uncensored · $0.97 / $1.94 | $6.402 | $8.342 | $14.162 |
Look at DeepSeek R1 0528 on Uncensored.com. Its input price is $0.50 per 1M tokens, but its output price is $2.18. With 4,000 thinking tokens, output makes up about 76% of the cost of each turn, and the turn costs 3.4 times as much as the same turn without thinking.
The listings for the newer DeepSeek builds don’t say whether they run in thinking mode by default. Check it on your first test requests: if completion_tokens in the usage object is much larger than the reply you got back, you’re paying for reasoning. Some APIs also report a separate reasoning token count. The pricing guide shows the full method for pricing a workload.
Which hosted DeepSeek models does the directory list?
This directory lists 4 DeepSeek-based models, from $0.12 to $0.97 per 1M input tokens. Here they are, cheapest input first:
| Provider | Model | Type | Input / output per 1M | Cached input | Context |
|---|---|---|---|---|---|
| Refuseless | DeepSeek V4.1 Flash 1017 Abliterated | Abliterated | $0.12 / $0.45 | $0.025 | 1M |
| Uncensored.com | DeepSeek R1 0528 | Standard model name | $0.50 / $2.18 | Not listed | Not listed |
| Uncensored.com | DeepSeek V4.1 Flash Cyber Uncensored | Uncensored build | $0.96 / $1.92 | Not listed | Not listed |
| UncensoredLLM | DeepSeek V4 Pro Uncensored | Uncensored build | $0.97 / $1.94 | Not listed | Not listed |
A few notes on reading it:
- The “Type” column only reflects the model name. A provider may serve a model without its own moderation filter even when the name doesn’t say “uncensored”, and a name that says “uncensored” doesn’t tell you how the model was modified. Ask the provider or test it.
- DeepSeek V4.1 Flash 1017 Abliterated on Refuseless also publishes a cached input rate of $0.025 per 1M tokens and a 1M context window. For chat apps that re-send long histories, caching cuts the input side of the bill further.
- NanoGPT also lists abliterated models (Abliterated Large Flash and Abliterated Model Large V2) whose names don’t give the base model. Check its model page before assuming either is DeepSeek-based.
- DeepSeek R1 0528 on Uncensored.com is the only full R1 in the list, under its plain name. Uncensored.com markets its models as unfiltered, but the listing doesn’t say whether this one was abliterated or fine-tuned, so test it on your own prompts.
For abliterated models on other base models, mostly GLM 5.3, see the best abliterated models comparison.
How to handle DeepSeek reasoning output in an API
Reasoning output comes back in one of two shapes, and your app has to handle both. Some OpenAI-compatible APIs return the thinking in a separate field on the message, often called reasoning_content. Others leave it inside the normal content string, wrapped in <think> and </think> tags. R1 distills write the tags into their raw output, and some local setups pass them straight through.
- Find where the thinking is. Send one request and log the full raw response. Check for a reasoning field and for
<think>tags in the content. - Hide it from users. Strip everything between the think tags before you display the reply, or show it in a collapsed panel if your users want it.
- Don’t send it back. Store only the final answer in the chat history. Re-sending old reasoning wastes input tokens on every later turn, and DeepSeek’s own API rejects plain chat requests that include
reasoning_contentin earlier messages. Tool calls can follow different rules, so check your provider’s documentation if you use them. - Raise
max_tokens. The limit covers thinking and answer together. A limit set for a 300-token reply can end the response mid-thought, with no visible answer at all. The SillyTavern setup guide covers the same issue for that client. - Plan for latency. Thinking is generated before the first visible token. Stream the response so users see activity, and set timeouts with that delay in mind.
DeepSeek’s model card for R1 also recommends a temperature around 0.6 and putting instructions in the user message rather than a system prompt. Abliterated distills inherit the same habits, so start from those settings and adjust from there.
Should you run DeepSeek R1 abliterated locally or use an API?
Run a distill locally if you have the hardware and want privacy or zero per-token cost. Use an API if you need a larger model, long context or high volume. The full R1 is out of reach for almost any single machine, so the real choice is between a local distill and a hosted model.
| Option | What you get | Main tradeoff |
|---|---|---|
| Local 7B or 8B abliterated distill | Runs on one consumer GPU or a recent laptop through Ollama. | Small model. Thin knowledge and weaker long-context handling. |
| Local 14B or 32B abliterated distill | Noticeably better reasoning. Quantized builds fit on a 24 GB GPU or a high-memory Mac. | Slow on modest hardware, and reasoning makes each reply longer. |
| Local 70B abliterated distill | The strongest distill. | Needs multiple GPUs or a large unified-memory machine. |
| Hosted DeepSeek build | Larger, newer models, long context, no hardware. | Per-token cost, provider terms and data handling to check. |
Local tokens aren’t free in practice either: thinking makes a reasoning model slower per reply, and a reply that takes a minute on your GPU may be a worse experience than a paid API. If you mainly want cheap uncensored chat, a non-reasoning model is often the better choice. The guide to free and cheap uncensored APIs compares local and hosted costs in more detail.
Checklist before you build on a DeepSeek abliterated model
- Identify the real base model. R1, an R1 distill (Qwen or Llama), or a newer DeepSeek release. Write down the exact model ID.
- Check the license. MIT for DeepSeek’s own weights, plus the Llama license for Llama-based distills.
- Measure thinking tokens. Compare
completion_tokenswith the visible reply on 20 to 50 real prompts, then price your workload with that number. - Count refusals in the reasoning too. Read a sample of the thinking as well as the final answer, and look for soft refusals and hedging.
- Test a non-reasoning model alongside it. For chat and roleplay, an uncensored GLM or DeepSeek Flash build without thinking may be cheaper, faster and good enough.
- Read the provider’s terms. Abliteration changes the model, not what the provider allows. Sexual content involving minors, and intimate images of real people made without consent, are illegal no matter which model produces them.
Frequently asked questions
What is DeepSeek R1 abliterated?
It usually means an abliterated version of one of the DeepSeek-R1-Distill models, such as the Qwen-32B or Llama-8B distill, with the refusal direction removed from the weights. The best-known builds come from huihui-ai on Hugging Face and Ollama. Abliterated copies of the full 671B R1 are much less common.
Is deepseek-r1-abliterated on Ollama the real DeepSeek R1?
No. The Ollama tags from 1.5b to 70b are distilled models: Qwen 2.5 or Llama 3 models fine-tuned on R1 outputs, then abliterated. They reason in R1’s style but are much smaller than the 671B R1.
Why are DeepSeek reasoning models more expensive than their price list suggests?
They write hidden reasoning before each answer, and most APIs bill that reasoning as output tokens. Output usually costs several times more than input, so a short visible reply can cost as much as a long one.
Which hosted DeepSeek abliterated model is cheapest?
In this directory, DeepSeek V4.1 Flash 1017 Abliterated on Refuseless is the cheapest DeepSeek-based model, at $0.12 per 1M input tokens and $0.45 per 1M output tokens. Prices change, so check the provider page before you build.
Can I use a DeepSeek abliterated model commercially?
DeepSeek’s R1 weights are MIT licensed, which allows commercial use and derivatives. Llama-based distills also carry Meta’s Llama license, and a hosted API adds its own terms, so check both for your use.
Sources
- DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning” (arXiv, 2025)
- DeepSeek-R1-0528 model card (Hugging Face)
- huihui-ai, DeepSeek-R1-Distill-Qwen-32B-abliterated model card (Hugging Face)
- huihui_ai/deepseek-r1-abliterated (Ollama)
- Arditi et al., “Refusal in Language Models Is Mediated by a Single Direction” (arXiv, 2024)