Qwen3.8 is a family of large language models developed by Alibaba Cloud. The series was introduced in 2025 as a successor to the Qwen2.5 generation, offering a range of parameter sizes from dense models to mixture-of-experts (MoE) architectures. The models are designed for both open-source distribution and commercial deployment through Alibaba Cloud's platform.
The Qwen3.8 family includes several configurations, with the flagship dense models at 8 billion and 32 billion parameters, alongside MoE variants such as the Qwen3.8-A14B, which activates 2.2 billion parameters per token. These models support a context length of 131,072 tokens and are trained on multilingual data covering over 30 languages, with particular strength in English and Chinese.
Architecture and Training
The Qwen3.8 models employ a standard Transformer (architecture) decoder-only architecture with enhancements such as multi-head attention, SwiGLU activation, and rotary position embeddings. The training process utilized a two-stage approach: initial pretraining on a large corpus of web text, books, and code, followed by supervised fine-tuning and reinforcement learning from AI feedback (RLAIF) to align with human preferences.
A notable feature of Qwen3.8 is its hybrid thinking mode, which allows the model to switch between rapid response generation and extended reasoning. This is controlled via a system prompt, enabling users to balance speed and depth of analysis. The models also incorporate greedy sampling and top-p sampling strategies during inference.
Performance and Benchmarks
On public benchmarks, Qwen3.8 models have demonstrated competitive performance. The Qwen3.8-32B model achieved a score of 75.1 on the MMLU-Pro benchmark, outperforming several larger models. On the AIME24 mathematics competition, the 32B model scored 67.0, while the MoE variant Qwen3.8-A14B scored 81.0. In coding tasks, the models recorded strong results on LiveCodeBench and SWE-bench, with the 32B model reaching a 45.0 pass rate on the latter.
The models have appeared on public LLM leaderboards, including the LMArena and Open LLM Leaderboard, where they have ranked among the top open-weight models in their size class. Independent evaluations have noted their efficiency, particularly the MoE variants which offer performance comparable to dense models at reduced computational cost.
Release and Availability
Qwen3.8 was released in stages, with the smaller 0.6B and 1.7B models made available first, followed by the 8B and 32B dense models, and finally the MoE variants. The weights are distributed under the Apache 2.0 license, permitting both research and commercial use. The models are available for download through Hugging Face and ModelScope, and can be deployed via Alibaba Cloud's Model Studio service.
Hardware support includes NVIDIA GPUs, with the models optimized for inference on AMD and Intel hardware as well. Deployment options range from local inference using vLLM or SGLang to cloud-based APIs. The release also included quantized versions using GPTQ and AWQ techniques for reduced memory footprint.
Reception and Impact
The Qwen3.8 series has been well-received in the open-source AI community, with developers praising its performance-to-size ratio and the flexibility of the hybrid thinking mode. The models have been integrated into various applications, from coding assistants to multilingual translation tools. The release has contributed to the broader trend of generative AI democratization, providing high-quality models that can run on consumer-grade hardware.
As of late 2025, the Qwen3.8 family remains actively maintained, with periodic updates and additional fine-tuned variants being released. The models have also served as a foundation for further research in areas such as model pruning and data augmentation.