Qwen3.8-Flash-Next Uncensored: Which Build to Pick
Qwen3.8-Flash-Next uncensored builds span 3 types. Compare abliterated weights, GGUF quants, and MTP builds to pick the right derivative for your hardware.
Key points
- Refusal-removed derivatives of the experimental Flash-Next appeared soon after release.
- They fall into three lines: edited base weights, requantized GGUF, and speed-focused MTP.
- Because the architecture is experimental, choose by your runtime support and use case.
Contents
Uncensored derivatives of Qwen3.8-Flash-Next fall into three groups: abliterated weights, GGUF re-quantizations and runtime builds, and speed-oriented Multi-Token Prediction (MTP) builds. Alibaba released the base model on August 24, 2026, as an experimental preview of the Qwen4-generation architecture. The first abliterated build, from huihui-ai, appeared two days later. Derivatives on Hugging Face have since passed 60, and 18 of the latest 60 search results carry refusal-removal terms such as abliterated or uncensored.
To run one locally with a tool like Ollama or LM Studio, you need a GGUF build, such as huihui-ai's GGUF or mradermacher's GGUF release. If you want faster generation and your runtime supports it, look at an MTP build. If you plan to quantize or fine-tune the model yourself, start from abliterated base weights. Qwen3.8-Flash-Next is a separate model from the dense Qwen3.8-27B, so the abliterated 27B editions are a different set of builds. Removing refusals does not make answers accurate, and the numbers publishers report are self-reported.
What is Qwen3.8-Flash-Next?#
Qwen3.8-Flash-Next is an experimental preview of the Qwen4-generation architecture, released by Alibaba's Qwen team on August 24, 2026. It is tagged image-text-to-text and licensed under qwen-community-1.0. It is a different model from the dense Qwen3.8-27B, which was released earlier, on August 5, 2026.
On Hugging Face, Qwen/Qwen3.8-Flash-Next had 1,299,909 downloads and 5,822 likes as of October 2, 2026. Its architecture highlights are Qwen Sparse Attention (QSA, a micro-block sparse attention), Gated Residual, N-gram Embedding, and a Muon/AdamW training recipe. The hosted cloud version, Qwen3.8-Flash, has a 1M-token context.
Qwen3.8-27B is an Apache 2.0 dense model with 27B parameters and a 262K context. If you want its abliterated editions, see Abliterated Qwen3.8-27B: Compare 4 Editions and Choose.
How many uncensored Qwen3.8-Flash-Next builds are there?#
Derivatives of Qwen3.8-Flash-Next on Hugging Face passed 60 within about five weeks of the August 24, 2026 release. The first abliterated build, huihui-ai's, appeared on August 26, two days after the base release. At least eight derivatives were published on a single day, October 2, and 18 of the latest 60 carry refusal-removal terms.
The 18 come from a Hugging Face search for "Qwen3.8-Flash-Next" sorted by creation date and checked on October 2, 2026. They contain terms such as uncensored, abliterated, heretic, obliterat or liberat. Because the search returns only 60 results, "more than 60" is the most that can be said about the total. Abliteration removes the refusal behavior from the weights themselves, as explained in What is abliteration?, and Uncensored vs abliterated: what's the difference? covers the naming. The derivatives fall into three groups: abliterated weights, GGUF re-quantizations and runtime builds, and speed-oriented MTP builds.
Which abliterated Qwen3.8-Flash-Next builds exist?#
The earliest abliterated Qwen3.8-Flash-Next build is huihui-ai's GGUF release of August 26, 2026, and mradermacher published a GGUF of it on October 2. Several other creators have published re-quantizations and runtime builds, including taurusduan, agentionai, SC117, wy2pay and jeffpeng3.
huihui-ai's huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated-GGUF is based on an image-text-to-text model, and its GGUF files come from Unsloth. The author used remove-refusals-with-transformers. The model card warns that safety filtering is greatly reduced and describes the release as research and experimental. For other models from this creator, see huihui-ai abliterated models.
mradermacher/Huihui-Qwen3.8-Flash-Next-abliterated-GGUF is mradermacher's GGUF version of the same build. Other community derivatives include taurusduan, agentionai (Gyro), SC117 (GSQ-RCO), wy2pay (IQ4XS), jeffpeng3 (EXL3 + Heretic) and Quat3rnion (halogen). Publishers of Flash-Next derivatives mostly do not report refusal rates or KL divergence, so there are no such numbers to compare here.
Which Qwen3.8-Flash-Next build should you choose?#
For running locally in a tool like Ollama or LM Studio, choose a GGUF build from huihui-ai or mradermacher. Choose a speed-oriented MTP build if faster generation matters and your runtime supports it. Choose abliterated base weights only if you plan to quantize or fine-tune the model yourself.
GGUF is the format local runtimes load, so a GGUF build is the practical default. How much VRAM it needs depends on the quantization level you pick, which GGUF quantization: how to choose Q4/Q8 explains.
What should you know before running one locally?#
Removing refusals does not make a Qwen3.8-Flash-Next build accurate, so verify important answers yourself. Numbers that publishers report are self-reported, and Flash-Next is an experimental preview. Also check how your interface handles thinking mode before you rely on the output.
An abliterated model declines fewer questions, but that does not add knowledge, so it can answer a sensitive question and still be wrong. Treat the huihui-ai build as the model card describes it, research and experimental.
What is Deepent Research?#
Deepent Research is a browser-accessible LLM chat service built to refuse as little as possible. It keeps going with research that a standard AI would cut off mid-answer.
- It refuses nothing. We run uncensored and abliterated models with guardrails reduced as far as possible, so prompts a standard AI answers with "I can't help with that" still get a response (Qwen3.5 / Qwen3.8-27B-abliterated, depending on your plan).
- Your input is never used for training. What you look up is used only to provide the service, never to train AI or for any other purpose. Personal data is protected, and our privacy policy is public.
- Try it free. No GPU and no install, just open a browser. In the logged-out trial your conversation isn't saved, so you can start casually.
→ Try it free, no sign-up: https://deepentresearch.com
No question can stop research.
Frequently asked questions#
Is Qwen3.8-Flash-Next abliterated an official release from Alibaba?#
No. Alibaba's Qwen team published the base Qwen3.8-Flash-Next on August 24, 2026, as an experimental preview under the qwen-community-1.0 license. The abliterated versions, such as huihui-ai's, are community modifications, and the GGUF conversions are made by community publishers as well. Alibaba's own hosted version is Qwen3.8-Flash, a separate cloud service.
How does Qwen3.8-Flash-Next differ from Qwen3.8-27B?#
Qwen3.8-Flash-Next, released on August 24, 2026, is an experimental preview of the Qwen4-generation architecture, with micro-block sparse attention and Gated Residual, under the qwen-community-1.0 license. Qwen3.8-27B, released on August 5, 2026, is a dense 27B model under Apache 2.0 with a 262K context. They are separate models, so abliterated 27B editions are not Flash-Next builds.
Which uncensored Flash-Next build should you download for local use?#
For local use in Ollama or LM Studio, download a GGUF build from huihui-ai or mradermacher, because local runtimes need GGUF. Pick a speed-oriented MTP build if faster generation matters and your runtime supports it. Use abliterated base weights only if you plan to quantize or fine-tune the model yourself. The quantization level you choose sets how much VRAM the build needs.
Does an abliterated Qwen3.8-Flash-Next guarantee accurate answers?#
No. Abliteration removes refusal behavior, so the model declines fewer questions, but it does not correct errors or add knowledge. An abliterated Qwen3.8-Flash-Next build can answer a sensitive question and still be wrong. Performance numbers on model cards are self-reported by the publisher, so verify important answers against reliable sources before you rely on them.
Disclaimer#
This article was drafted with AI assistance and then reviewed, fact-checked, and edited by the Deepent Research team before publication.
It is provided for educational and research purposes and is not intended to encourage unlawful activity or crime.
Where figures are cited, they come from third-party public reports and vendor-published values and may vary with measurement conditions.
How this note was edited and verified
We cross-check primary sources such as official documentation and papers, and organize only what the facts support.
This note was drafted with AI, then fact-checked, expanded, and edited by the operator before publication.
Related notes
Abliterated Qwen3.8-27B: Compare 4 Editions and Choose
Abliterated Qwen3.8-27B differs by creator. Compare huihui-ai, orcarouter, Heretic and RVN editions by edit range, quantization and published refusal rates.
What Is huihui-ai (huihui AI)? Abliterated Models Guide
huihui-ai (huihui AI) publishes abliterated versions of open models on Hugging Face. See popular builds, how to read model names, and which size to pick.
Uncensored LLMs: 10 Best Models by GPU Memory (2026)
In 2026, the best uncensored LLM for most people is Qwen3.8-27B abliterated. 10 top picks by GPU memory, plus how uncensored and abliterated models differ.
Try it yourself
Deepent Research is an LLM chat service that refuses almost nothing. Everything covered in this note can be tried without downloading anything.