Alibaba Unveils Qwen3.8-Flash-Next AI Model with 176B Parameters

Alibaba has released the weights for its Qwen3.8-Flash-Next multimodal model, a 176-billion parameter Mixture-of-Experts (MoE) architecture designed to handle long-context tasks such as agentic coding, document processing, and high-volume workflows. This release, announced on August 26, serves as a technical preview of the upcoming Qwen4 model family and emphasizes cost-efficiency and scalability for developers.

Technical Highlights

Qwen3.8-Flash-Next combines two key innovations for long-context inference: Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA). GDN compresses historical context into a fixed-size recurrent state, avoiding the memory growth typically associated with long sequences. Meanwhile, QSA works at the micro-block level to reduce computational overhead while maintaining retrieval precision. Together, these mechanisms enable efficient processing of sequences up to 1 million tokens, an area where traditional attention models struggle.

Benchmarks shared by Alibaba indicate QSA delivers tangible performance gains. For example, in a 1-million-token workload with a 90% prefix-cache hit rate, Qwen3.8 achieved 8.6x the prefill throughput of its predecessor, Qwen3.7-Plus. During prefill and decoding phases, the sparse attention kernel outperformed full attention by 7.6x and 4.9x, respectively, underscoring its utility for high-load applications.

Optimized for NVIDIA GB300 NVL72

NVIDIA has collaborated with Alibaba to validate Qwen3.8-Flash-Next on its GB300 NVL72 platform, which integrates 72 Blackwell Ultra GPUs with a 130 TB/s NVLink communication fabric. The model achieves over 16,000 tokens per second per GPU and supports more than 200 tokens per second per user, making it suitable for large-scale deployment in production environments. Developers can also experiment with the model on smaller systems like NVIDIA DGX Spark clusters or workstations equipped with RTX PRO 6000 Blackwell GPUs.

Developer Tools and Accessibility

Qwen3.8-Flash-Next is available as open weights, and developers can download it from platforms like Hugging Face and ModelScope. NVIDIA supports fine-tuning through its NeMo AutoModel library, enabling domain-specific adaptations via efficient methods such as LoRA. Reinforcement learning workflows are also supported using NeMo RL recipes.

For inference, developers can choose from multiple tools, including SGLang, vLLM, and NVIDIA TensorRT, ensuring flexibility in deployment. Alibaba has positioned this release as a cost-efficient model for developers to prototype and scale workflows ahead of the Qwen4 launch.

Why This Matters

Qwen3.8-Flash-Next represents a significant milestone in the evolution of large language models. By addressing the scalability challenges of long-context tasks, Alibaba is laying the groundwork for broader adoption of multimodal AI solutions. While this release is primarily a technical preview, its innovations in sparse attention and memory efficiency could influence the next generation of AI architectures and tools.