Qwen3.8-27B Derivative Models Flood Hugging Face Trends
The cluster shows how permissive open-weight licensing can turn a brand-new model into a wave of modified releases, including safety-removed variants, within days.
Key Facts
- As of August 22, 17 of the top 30 trending models on Hugging Face were Qwen3.8-27B entries, with 15 being derivative versions.
- Ten of the 17 Qwen3.8-27B models were uncensored versions, with names including Uncensored, abliterated, OBLITERATED, and Heretic.
- The 10 uncensored Qwen3.8-27B models totaled approximately 2.78 million downloads.
- Qwen3.8-27B is a 27-billion-parameter multimodal model released as open weights on August 15 under the Apache 2.0 license.
- Unsloth's GGUF version of Qwen3.8-27B had 5.8 million downloads over 30 days, while the official repository had 1.73 million and the FP8 version 1.94 million.
Reporting from 2 sources: GIGAZINE, GameBusiness.jp.
As of August 22, Hugging Face model trends were dominated by derivatives of Qwen3.8-27B, Alibaba's Qwen team's open-weights multimodal model. Of the top 30 trending models, 17 were Qwen3.8-27B entries; excluding the two official Qwen repositories, 15 were derivative versions. Ten of those 17 were uncensored versions, with names containing terms like Uncensored, abliterated, OBLITERATED, and Heretic, meaning their safety mechanisms were removed. The combined downloads for those 10 uncensored models reached approximately 2.78 million. Qwen3.8-27B, released on August 15, has 27 billion parameters, can run on a single high-end GPU, and is licensed under Apache 2.0, which permits free modification and redistribution. That license also does not prohibit uncensored derivatives. The popularity extends beyond the uncensored set: Unsloth's GGUF version logged 5.8 million downloads over the past 30 days, while the official repository saw 1.73 million downloads, plus 1.94 million for the FP8 version. The flood arrived within roughly a week of the model's release.
- Unsloth's GGUF version of Qwen3.8-27B leads all derivatives with 5.8 million downloads over the past 30 days.
- Official Qwen repository logged 1.73 million downloads, plus 1.94 million for its FP8 version.
- Ten uncensored models with names like Uncensored, abliterated, OBLITERATED, and Heretic account for roughly 2.78 million combined downloads.
- Apache 2.0 licensing does not prohibit uncensored derivative releases, which is why they appeared within about a week.
- Qwen3.8-Flash-Next, released August 26 on Hugging Face and ModelScope, uses a hybrid architecture with Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA), and its training time is one-ninth that of Qwen3.8-27B.
- Qwen3.8-Flash-Next specs: 125 billion total parameters, 6 billion active, 51 billion N-gram Embedding parameters, 262,144-token standard context extendable to 1 million via YaRN.
- Unsloth's quantized Flash-Next released the same day; the 1-bit version runs on machines with over 75GB of RAM and maintains 80% accuracy of the full model.
- Benchmark note: Qwen3.8-Flash-Next outperforms the proprietary (Max) model in many tests and is 8.6 times faster than Qwen3.7-Plus when processing 1 million tokens.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.