What are DeepSeek Uncensored models?
A comparison of a 510 GB DeepSeek weight edit and a compact Qwen student model: lineage, refusal tests, hosted prices, local hardware and evidence limits.
LLM Pricing2,115 words8 min read
On this page
DeepSeek Uncensored names several community derivatives, none released by DeepSeek. The two recent lines most likely to be confused have different foundations: DeepSeek V4.1 Flash 1017 Uncensored edits the weights of DeepSeek's 552B mixture-of-experts model, while DeepSeek V4 Pro Qwen3.5 9B Uncensored Heretic is a Qwen3.5-9B student trained on answers attributed to DeepSeek V4 Pro and edited again to reduce refusals.
Short answer
- DeepSeek has not published an official model called DeepSeek Uncensored. Each listing needs a specific base model, weight author and revision.
- DeepSeek V4.1 Flash 1017 Uncensored is a weight-edited version of DeepSeek V4.1 Flash 1017. Its upstream checkpoint supports a 1M-token context, images, video, tools and reasoning tiers, but its roughly 510 GB weight set was validated on four H200 GPUs.
- DeepSeek V4 Pro Qwen3.5 9B Uncensored Heretic uses Qwen3.5 9B as its architecture. DeepSeek V4 Pro supplied the claimed teacher data; the resulting model is not a smaller copy of DeepSeek V4 Pro.
- Publisher tests report far fewer refusals for both lines. Only the V4.1 releases include a before-and-after task benchmark; the 9B Heretic release reports refusals and KL divergence, not math or general-knowledge accuracy after the edit.
- The official V4.1 Flash 1017 API costs $0.15 input, $0.60 output and $0.003 cached input per 1M tokens. A hosted uncensored endpoint was $0.33 / $1.32 / $0.01 on Nov 9, 2026; the 9B endpoint seen by this site on Nov 8 had disappeared from the next catalog snapshot.
One label, two model lineages
The full-size line starts with DeepSeek. DeepSeek released V4.1 Flash on Sep 10, 2026, then replaced the API model with the 1017 revision on Oct 17. Community authors modified both checkpoints at the weight level. The first V4.1 uncensored repository checked for this article was created on Sep 10; the first 1017 repository appeared on Oct 17, followed by dealignai's FP8 release on Oct 18.
The 9B line starts with Qwen. Jackrong fine-tuned Qwen3.5-9B using about 250,000 mathematics and STEM examples that the publisher says were generated by DeepSeek V4 Pro in Max Effect mode. The training description lists supervised fine-tuning followed by a light GSPO reinforcement-learning stage. xtools-at then applied Heretic v2 Arbitrary-Rank Ablation to that checkpoint and published the uncensored weights on Oct 30.
Those paths can be written without the ambiguous product names:
- DeepSeek V4.1 Flash 1017 was weight-edited into the full-size uncensored model.
- Qwen3.5-9B was fine-tuned on claimed DeepSeek V4 Pro outputs, then weight-edited into the compact Heretic model.
DeepSeek is the base-model maker in the first path and the claimed teacher in the second. The shared name records provenance and the intended reduction in refusals; it does not establish a common architecture or level of capability.
The full V4.1 Flash 1017 build
DeepSeek V4.1 Flash 1017 is a 552B-parameter multimodal mixture-of-experts model with 8B parameters active during prefill and 16B during decoding, according to the derivative model cards. The official checkpoint supports a 1M-token context, image and video input, tool calling and selectable reasoning effort. DeepSeek's Oct 17 update reported gains over the September release on coding, agent and visual benchmarks, including Terminal-Bench 4.0 rising from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0.
The dealignai uncensored release keeps the native FP8 and FP4 layout and changes selected weight tensors rather than adding a system prompt, adapter or runtime hook. Its repository contains 48 weight shards and about 510 GB of stored files. The author validated it on four H200 GPUs; the model card reports about 564 GB of total VRAM for that machine and notes that part of the model's Engram memory can be moved to host RAM. The upstream 1M context is therefore a model limit, not a promise that every hosted endpoint exposes the full window.
A separate 1017 build by mrfakename used a rank-2 refusal subspace because chat and thinking modes showed different refusal directions. That release reported 4 refusals in 100 chat prompts and 9 in 100 thinking prompts. The difference from dealignai's results matters: “Uncensored” describes a goal shared by several edits, not one interchangeable checkpoint.
The compact Qwen3.5 9B Heretic build
The compact model keeps Qwen3.5's 32-layer hybrid text architecture and vision encoder. Its configuration declares a 262,144-token maximum position length, image and video tokens, thinking output and a tool-call template. The BF16 repository holds about 19.3 GB of files and uses the Apache 2.0 licence. Its DeepSeek V4 Pro name refers to the training-data lineage reported by Jackrong.
The original distilled checkpoint has publisher-reported results of 94.50% on GSM8K and a 90.53% average across 500 sampled questions from each of the MMLU-Pro mathematics, physics and chemistry subsets. The same card reports 89.60% for unmodified Qwen3.5-9B on that three-subset comparison. These numbers were measured on different quantized builds and backends, and they belong to the checkpoint before xtools-at applied Heretic.
Local conversions make the size difference tangible. One Q4_K_M release is a 5.24 GiB text model with a separate 876 MiB vision projector. An NVFP4 conversion loaded its text weights in 9.65 GiB on a 16 GB RTX Pro 2000 Blackwell and generated about 29 tokens per second in the publisher's single-request test at a 16K server context. Native quantized vision serving was not tested in that run. These measurements describe those files and settings; they are not hardware requirements for every conversion.
What the refusal tests establish
The published tests do not use one shared prompt set or grader, so their refusal counts cannot rank the models against each other.
| Release | Publisher's refusal result | Capability check after editing | Main limit |
|---|---|---|---|
| dealignai V4.1 Flash 1017 Uncensored | 320/320 classified as compliant at maximum reasoning effort | 83.77% versus 87.63% on a 1,140-question MMLU sample, −3.86 percentage points | Base used the official API; edited model used a self-hosted engine |
| mrfakename V4.1 Flash 1017 Abliterated | 4/100 refusals in chat; 9/100 in thinking | KL divergence 0.067 on harmless first-token distributions | Reasoning traces were judged before some answers finished |
| xtools-at Qwen3.5 9B Heretic | 3/100 refusals versus 100/100 for its source checkpoint | KL divergence 0.0422 | No post-edit task-accuracy benchmark is published |
KL divergence measures how far the edited model's output distribution moved on the chosen benign prompts. A small value supports a narrow claim about distribution shift under that test. It does not show that GSM8K, MMLU-Pro, vision or tool calling kept the same accuracy. The 9B model's inherited math scores therefore remain background about its source checkpoint, not measured scores for the uncensored release.
The V4.1 results also show a real trade-off. dealignai's earlier September build reported MMLU falling from 86.96% to 82.74%, with most of the loss concentrated in moral-scenario and professional-law questions. Its 1017 test found a smaller decline on a sampled set. Removing refusals can leave many ordinary tasks close to the base while still changing the answers used by benchmarks that reward hesitation or refusal.
API price and local size
Prices are USD per 1M tokens, checked on Nov 9, 2026. “Historical” means the endpoint appeared in a saved catalog snapshot and was absent from the next one; it does not prove that every provider stopped serving the weights.
| Model or endpoint | Status | Input | Output | Cache read | Exposed context |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 official | Current official API | $0.15 | $0.60 | $0.003 | 1M |
| V4.1 Flash 1017 Uncensored, Venice | Current when checked | $0.33 | $1.32 | $0.01 | Provider-dependent, up to the 1M model limit |
| V4 Pro Qwen3.5 9B Uncensored, NanoGPT | Historical, Nov 8 | $0.28 | $0.28 | $0.14 | 32K |
| R1 Distill Llama 70B Abliterated, NanoGPT | Current when checked | $0.70 | $0.70 | $0.35 | 16.4K |
| R1 Distill Qwen 32B Abliterated, NanoGPT | Current when checked | $1.40 | $1.40 | $0.70 | 16.4K |
At the checked rates, the hosted full-size uncensored build cost 2.2 times the official 1017 base for uncached input and output. The compact 9B listing was cheaper per output token, but it was a different model and disappeared after one snapshot. Comparing the two prices as if they bought the same capability would ignore the Qwen architecture, the distillation step, the much smaller deployment and the missing post-edit task evaluation.
The older R1 listings provide another warning about names. One derives from Llama 3.3 70B and the other from Qwen2.5 32B; both were distilled from DeepSeek R1 responses before their weights were edited. Their current hosted limits are smaller than the source checkpoints' published limits, and neither community repository provides a before-and-after task benchmark.
How the choices differ
The full V4.1 Flash 1017 build preserves the architecture, multimodal inputs and agent features of a recent DeepSeek checkpoint. Its scale makes hosted inference or data-centre hardware the practical route, and its publisher provides the clearest before-and-after evidence of capability loss. It also costs more per token than DeepSeek's official endpoint.
The 9B Heretic build fits a different use case: local experiments with a compact multimodal Qwen model carrying a specific DeepSeek-derived STEM fine-tune. Consumer-GPU conversions exist, yet their quantization, context allocation, vision support and MTP handling vary. The release supplies a refusal test but no repeated math, knowledge or tool benchmark after the Heretic edit.
The R1 Distill variants remain available through one catalogued host and start from older, text-only checkpoints. Their model cards describe the abliteration as a proof of concept. Availability alone supplies less evidence than a measured comparison with the unedited weights.
What to keep in mind
- Check the complete repository ID. “DeepSeek V4 Pro 9B” can refer to a Qwen student model, while “V4.1 Flash 1017 Uncensored” can refer to several independently edited DeepSeek checkpoints.
- Keep model limits separate from provider limits. The 9B weights declare 262K context, but the former hosted endpoint exposed 32K; V4.1 supports 1M upstream, while a host may set a lower window.
- Treat publisher benchmarks as publisher benchmarks. The available tests are useful evidence, but no independent evaluation has compared all these uncensored releases under one protocol.
- Reduced refusals remove a layer of protection. A deployment that needs input or output moderation must add and test it separately.
- Verify the exact quantized file. File size, memory use, vision support and speculative decoding can change between BF16, NVFP4 and GGUF releases even when their names share the same model stem.
Frequently asked questions
Is DeepSeek Uncensored an official DeepSeek model?
No. DeepSeek publishes the base V4.1 Flash and V4 Pro models. The uncensored versions covered here were produced by community authors who modified open weights or a community fine-tune.
Is DeepSeek V4 Pro Qwen3.5 9B a small DeepSeek V4 Pro?
Its architecture is Qwen3.5-9B. The publisher says DeepSeek V4 Pro generated the mathematics and STEM answers used to fine-tune it, so the DeepSeek name records its teacher lineage rather than its architecture.
Which DeepSeek Uncensored model can run on a consumer GPU?
The Qwen3.5 9B Heretic line has consumer-sized conversions. A tested NVFP4 text deployment loaded in 9.65 GiB on a 16 GB GPU, while a Q4_K_M text file is about 5.24 GiB. The full V4.1 Flash checkpoint is about 510 GB and was validated on four H200 GPUs.
Does uncensored mean the model answers every request?
No shared specification guarantees that. Published results range from zero classified refusals in one V4.1 test to 3, 4 or 9 refusals per 100 in other tests, depending on the checkpoint, prompt set, reasoning mode and grader.
Did the 9B Heretic model keep the original math scores?
That has not been demonstrated in the published material. The 94.50% GSM8K and 90.53% three-subset MMLU-Pro figures belong to Jackrong's source checkpoint. xtools-at reports refusal count and KL divergence after editing, without a post-edit task benchmark.
How much does DeepSeek Uncensored cost?
The checked V4.1 Flash 1017 Uncensored endpoint charged $0.33 input and $1.32 output per 1M tokens. The Qwen3.5 9B endpoint briefly seen by this site charged $0.28 for either input or output on Nov 8, then disappeared from the next snapshot. Prices and model revisions need to be checked together.
What licences apply?
DeepSeek V4.1 Flash 1017 and its cited uncensored FP8 derivative use MIT. Qwen3.5 9B, the Jackrong fine-tune and the xtools-at Heretic repository use Apache 2.0. A converted file should retain the notices required by its source licence.
Sources
Read on Nov 9, 2026. Prices are USD per 1M tokens. Benchmark figures are publisher-reported unless stated otherwise.
- DeepSeek: V4.1 Flash 1017 update
- DeepSeek: DeepSeek-V4.1-Flash-1017 model card
- dealignai: DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8
- dealignai: DeepSeek-V4.1-Flash-UNCENSORED-FP8
- mrfakename: DeepSeek-V4.1-Flash-1017-abliterated
- Venice: DeepSeek V4.1 Flash 1017 Uncensored pricing
- Qwen: Qwen3.5-9B model card
- Jackrong: DeepSeek-V4-Pro-Qwen3.5-9B
- xtools-at: DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic
- sakamakismile: NVFP4 conversion and hardware test
- Sakura721: Q4_K_M conversion and file sizes
- LLM Pricing: Nov 8 catalog snapshot
- huihui-ai: DeepSeek R1 Distill Llama 70B Abliterated
- huihui-ai: DeepSeek R1 Distill Qwen 32B Abliterated
Related articles
All articlesWhat are GLM 5.3 Uncensored and GLM 5.3 Flash Uncensored?
Two weight-edited GLM-5.3 derivatives that refuse far less than the base models. What abliteration changes, which base each starts from, what each costs, and the limits that still apply.
1,324 words5 min read- ExplainerAI models
What is Jev? TypeSafe's decision model explained
A model that picks instead of writing. What a System One decision model takes in and returns, what it costs, how independent tests rate it, and where it breaks.
2,579 words10 min read