Alibaba Releases Qwen3.8-Flash with Innovative Model Architecture Delivering Optimal Price-Performance
Alibaba has unveiled Qwen3.8-Flash, an open-weight, multimodal Mixture-of-Experts (MoE) model that delivers exceptional value for money. Qwen3.8-Flash features a 125B-parameter main model, supplemented by an additional 51B N-gram embeddings, with 6B parameters activated per token. It strikes an optimal balance between capability, latency, and cost — making it ideal for high-volume applications, tool-driven workflows, and coding or co-working assistants. The model also serves as an early preview of the architecture that is designed to power the upcoming Qwen4 series.
Qwen3.8-Flash demonstrates remarkable capabilities in agentic coding, long-horizon agent tasks, and multimodal intelligence. It performs competitively against leading models such as DeepSeek-V4-Flash and Claude-Opus-4.6 across multiple benchmarks, including SWE-bench Pro (agentic coding), CoWorkBench (long-horizon office work), Toolathlon Verified (real-world tool use), MathVision (visual math problem solving), AndroidWorld (agentic mobile use) and ERQA (embodied intelligence).
The model natively supports 262K tokens of context and can be extended to 1 million tokens. Its weights are now available on Hugging Face and ModelScope for global developers to download and use. The model can be accessed via API at competitive rates on Model Studio and Qwen Cloud, Alibaba’s AI-native cloud platform. The pricing per 1 million tokens is 1 RMB (0.16 USD) for input and 3 RMB (0.47 USD) for output.
The model is also available on QwenWork, Alibaba’s all-in-one workplace AI agent platform, where it powers a redesigned Standard mode that cuts token consumption per task by 75% and roughly doubles generation speed of the current mode, bringing flagship-level capability within reach of everyday workloads.
Architectural Innovations Drive Compute Efficiency
Qwen3.8-Flash introduces architectural innovations across attention mechanisms, residual connections, embeddings, and optimization. As a result, it enhances model capability while further improving computational efficiency, capacity, and training stability. For instance, its hybrid attention architecture combines Gated DeltaNet (GDN), which efficiently compresses historical information, with Qwen Sparse Attention (QSA), a novel design that uses a lightweight compressed indexer to select relevant context, substantially reducing attention costs for long sequences. Additionally, the Gated Residual (GR) mechanism expands data pathways between layers while strengthening cross-layer information flow and training stability. The N-gram Embedding technique scales model capacity with minimal additional computation, and the Muon Optimizer further enhances large-scale model training efficiency.
Driven by architectural innovations, Qwen3.8-Flash significantly reduces both training and inference costs compared to Qwen3.7-Plus—a model three times its size. It requires only about one-ninth of the training resources while delivering superior performance in coding and office tasks.
Across the open-source community including both Hugging Face and ModelScope, Alibaba has open-sourced more than 460 models. This ecosystem has spawned over 300,000 derivative models and accumulated more than 3 billion global downloads, making it the world’s most-downloaded open-source model family.
