# AI-Powered Coding with Qwen3.8-Flash-Next High-performance AI experiments on NVIDIA GB300 NVL72 - **Source:** [Nvidia Developer Blog](https://developer.nvidia.com/blog) - **Published:** 2026-08-26 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/7105a890-9dd0-46d8-b351-3c092afc26e2 ## Technical Specifications of Qwen3.8-Flash-Next Qwen3.8-Flash-Next is a 125-billion parameter multimodal mixture-of-experts (MoE) model. It activates 6 billion parameters per token and features a native 262,144-token context window, expandable to 1 million tokens with YaRN. The model uses a hybrid architecture combining Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA), optimized for large-context applications. ## Advantages of Hybrid Architecture Three out of four layers use GDN to compress historical context into a fixed-size recurrent state, eliminating KV cache growth as sequences lengthen. The remaining layer employs QSA for precise retrieval across full context. QSA aggregates sequences into micro-blocks, estimates their importance at block level, and selects only the most relevant regions. ## Performance Improvements Alibaba benchmarks show QSA improves efficiency for 1M-token workloads. Compared to full attention, it delivers up to 7.6x speedup during prefill and 4.9x during decoding. At 1M-token context length with 90% prefix-cache hit rate, Qwen3.8-Flash-Next achieves 8.6x the prefill throughput of Qwen3.7-Plus. ## Running on NVIDIA GB300 NVL72 The NVIDIA GB300 NVL72 features a rack-scale architecture integrating 72 Blackwell Ultra GPUs. Its large 130 TB/s NVLink domain eliminates bottlenecks from traditional networks. On this platform, Qwen3.8-Flash-Next processes over 16,000 tokens per second per GPU. ## Local Hardware Support The model also runs on local NVIDIA hardware including DGX Station, DGX Spark clusters, and workstations with four RTX PRO 6000 Blackwell Max-Q GPUs. Developers can prototype on local hardware and scale the same model to GB300 NVL72 for production. ## Fine-tuning and Learning Options NVIDIA NeMo AutoModel offers PyTorch-native fine-tuning with Day-0 Hugging Face checkpoint support. Options include full SFT or memory-efficient LoRA fine-tuning. Reinforcement learning is available through NVIDIA NeMo RL recipes. ## Inference Engine Options NVIDIA supports multiple inference stacks for diverse developer needs. SGLang, vLLM, and TokenSpeed provide open-source inference recipes for greater performance control on NVIDIA-accelerated platforms. ## The Future of AI-Powered Coding Qwen3.8-Flash-Next's large context window and hybrid architecture usher in a new era for AI-assisted coding. Its high throughput and low latency on platforms like NVIDIA GB300 NVL72 hold revolutionary potential for complex software development processes. ## Sources covering this story - https://developer.nvidia.com/blog/experiment-with-qwen3-8-flash-next-on-nvidia-gb300-nvl72-for-agentic-coding/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._