Qwen3.8-Flash-Next Debuts on NVIDIA GB300 NVL72 for Long-Context AI

Alibaba has released the model weights for Qwen3.8-Flash-Next, an experimental multimodal Mixture-of-Experts (MoE) AI model designed to push the boundaries of long-context processing. Positioned as a preview of the upcoming Qwen4 architecture, the model boasts cutting-edge innovations aimed at high-efficiency, high-context applications. Developers can now run Qwen3.8-Flash-Next on NVIDIA’s powerful GB300 NVL72 platform, achieving unprecedented throughput and scalability for agentic coding, document processing, and more.

The Qwen3.8-Flash-Next model features a 125-billion parameter base, augmented by 51 billion N-gram embeddings, and a native 262,144-token context window extensible to 1 million tokens using the YaRN framework. This makes it particularly suited for tasks requiring extensive sequential data processing, such as legal document analysis or tool-driven automation workflows.

Technical Innovations: GDN and QSA

At the core of Qwen3.8-Flash-Next are two architectural breakthroughs: Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA). GDN eliminates the memory burden of growing key-value (KV) caches by compressing historical context into a fixed-size recurrent state. This ensures consistent performance regardless of sequence length. Meanwhile, QSA addresses computational inefficiencies in long-context attention by aggregating sequences into micro-blocks and retrieving only the most relevant regions, significantly reducing overhead.

Benchmarks published by Alibaba validate these innovations. At a 1 million-token context length, the QSA attention kernel reportedly delivered 7.6x speedups in prefill and 4.9x in decoding compared to traditional full-attention mechanisms. In a real-world serving test with a 90% prefix-cache hit rate, Qwen3.8-Flash-Next achieved 8.6x prefill throughput over its predecessor, Qwen3.7-Plus.

Performance on NVIDIA GB300 NVL72

The NVIDIA GB300 NVL72 platform, featuring 72 Blackwell Ultra GPUs and a 130 TB/s NVLink communication domain, is the backbone for Qwen3.8-Flash-Next’s deployment. This setup enables the model to process over 16,000 tokens per second per GPU, equating to more than 200 tokens per second per user. Such performance highlights its potential for high-throughput, low-latency applications.

For developers, the scalability of Qwen3.8-Flash-Next is noteworthy. While large-scale deployments can leverage the GB300 NVL72, smaller setups can use NVIDIA DGX Station or workstations equipped with RTX PRO 6000 GPUs for prototyping.

Fine-Tuning and Deployment

Developers aiming to customize the model for specific use cases can take advantage of NVIDIA NeMo tools. The NeMo AutoModel library supports fine-tuning directly on the provided checkpoints, either via full supervised fine-tuning (SFT) or memory-efficient low-rank adaptation (LoRA). For reinforcement learning, NVIDIA’s NeMo RL recipes offer additional optimization pathways.

Inference options are flexible, with supported stacks including SGLang, vLLM, and TokenSpeed. Open-weight availability on Hugging Face and ModelScope ensures accessible experimentation.

Looking Ahead

Qwen3.8-Flash-Next is a stepping stone toward the Qwen4 model family, with Alibaba using it as a testbed for next-generation architectural concepts. Its innovations in attention mechanisms and memory efficiency suggest transformative potential for long-context AI applications. Developers now have the opportunity to experiment with this cutting-edge model and prepare for the Qwen4 release.

Image source: Shutterstock