Explore additional models
Frequently
asked questions
Qwen3.8 27B FFT is the full fine-tuning version of Qwen3.8 27B, the dense model the Alibaba team released in August 2026 under the Apache 2.0 license. Qwen3.8 builds on the architecture of Qwen3.5. On Crusoe, Qwen3.8 27B FFT is available for Serverless Fine-Tuning in Crusoe Intelligence Foundry.
FFT stands for full fine-tuning. A full fine-tuning job trains on your dataset and updates every weight in Qwen3.8 27B, across the attention and feed-forward blocks of all 64 layers. Teams typically choose full fine-tuning for larger datasets or broad changes, such as adapting the model to a new domain or a private codebase.
No. Qwen3.8 27B is a dense model, so all 27B parameters run on every token. Alibaba also trained it with a multi-token prediction head, which inference engines can use for speculative decoding.
The Alibaba team built Qwen3.8 for coding, professional work, research, and long-horizon agentic tasks. They designed Qwen3.8 27B to carry complex, multi-step tasks through to completion. Teams can fine-tune it on their own training data for that kind of work. Examples include coding in a private codebase; professional work in fields like computer science, finance, law, and medicine; and research workflows that follow internal formats and terminology.


%204.png)
