Experimenting With the Qwen3.8-Flash-Next 176B Model on NVIDIA GB300 NVL72

Alibaba has released the model weights for Qwen3.8-Flash-Next as a preview of its upcoming Qwen4 architecture. The release is intended for developer experimentation and evaluation.

Qwen3.8-Flash-Next is a multimodal mixture-of-experts (MoE) model with 176B total parameters. This includes 51B N-gram embedding parameters, while 6B parameters are activated for each token. The model provides a native 262,144-token context window, which can be extended to 1M tokens with YaRN.

NVIDIA provides best-effort Day 0 functional support through SGLang, vLLM, and NVIDIA TensorRT LLM. The model has also been validated for inference on NVIDIA GB300 NVL72, with post-training recipes available through NVIDIA NeMo AutoModel and NVIDIA NeMo RL.

Architectural Innovations for Long-Context Inference

Qwen3.8-Flash-Next targets high-volume, context-intensive workloads such as agentic coding, document processing, and tool-driven workflows. As context windows expand, attention computation and key-value (KV) cache memory can become bottlenecks.

The model addresses these issues with a hybrid architecture that combines Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA). Three out of every four layers use GDN, which continuously compresses historical context into a fixed-size recurrent state. This prevents KV cache requirements from increasing as sequences become longer. The remaining layer uses QSA to retrieve specific information from the full context.

Earlier sparse-attention approaches often use token-level indexers, whose computational cost rises with context length. QSA instead divides the sequence into micro-blocks, estimates the importance of each block, and selects the most relevant regions. This reduces attention, computation, and indexing overhead within each layer, complementing the alternating GDN and QSA design.

Alibaba's published benchmarks indicate that QSA can improve the efficiency of 1M-token workloads. Compared with full attention, its attention kernel achieved speedups of up to 7.6x during prefill and 4.9x during decoding. In a cache-heavy online-serving test with a 1M-token context length and a 90% prefix-cache hit rate, Qwen3.8-Flash-Next delivered 8.6x the prefill throughput of Qwen3.7-Plus.

Running Qwen3.8-Flash-Next on NVIDIA GB300 NVL72

The GB300 NVL72 uses a rack-scale architecture that integrates 72 NVIDIA Blackwell Ultra GPUs into one platform. Its 72-GPU NVIDIA NVLink domain supports all-to-all communication at 130 TB/s. This helps avoid bottlenecks that can occur when expert traffic must pass through conventional, off-the-shelf networks.

On NVIDIA GB300 NVL72, Qwen3.8-Flash-Next delivers more than 16K tokens per second per GPU and more than 200 tokens per second per user. These performance levels allow developers to test agentic coding applications with high throughput and low latency.

The model can also run on local NVIDIA hardware, including NVIDIA DGX Station, NVIDIA DGX Spark clusters, and workstations equipped with four NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPUs. Developers can prototype and evaluate agentic coding workflows locally, then use the same model on GB300 NVL72 for production serving.

Post-Training and Inference Options

Developers can fine-tune Qwen3.8-Flash-Next for domain-specific applications with NVIDIA NeMo AutoModel, a PyTorch-native fine-tuning library that provides Day 0 support for Hugging Face checkpoints. Training can begin directly from existing checkpoints without model conversion. The library supports full supervised fine-tuning (SFT) and memory-efficient LoRA fine-tuning.

For reinforcement learning, developers can use the available NVIDIA NeMo RL recipes.

Several inference stacks are supported for different deployment requirements. SGLang, vLLM, and TokenSpeed provide open-source inference recipes for developers who need control over performance on NVIDIA-accelerated systems.

Getting Started

The model can be tried through QwenCloud. Its weights are available for download from Hugging Face and ModelScope.