Alibaba released the model weights for Qwen3.8-Flash-Next as a preview of the upcoming Qwen4 architecture for developers to experiment with and evaluate. It’s a multimodal mixture-of-experts (MoE) model with a 125B-parameter main model supplemented by an additional 51B N-gram embeddings, with 6B parameters activated per token. It has a native 262,144-token context window, extensible to 1M tokens with YaRN.
NVIDIA provides best-effort Day 0 functional support through SGLang, vLLM, and NVIDIA TensorRT LLM, validation across NVIDIA GB300 NVL72 for inference, and post-training recipes from NVIDIA NeMo AutoModel and NVIDIA NeMo RL.
Architectural innovations for long-context inference #
Qwen3.8-Flash-Next is designed for high-volume, context-intensive applications such as agentic coding, document processing, and tool-driven workflows. As context grows, attention compute and KV cache memory become bottlenecks. The model addresses both with a hybrid architecture combining Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA). Three out of every four layers use GDN to continuously compress historical context into a fixed-size recurrent state, eliminating KV cache growth as sequences lengthen. The remaining layer uses **QSA **for precise retrieval across the full context.
Previous sparse-attention approaches rely on token-level indexers that become increasingly computationally expensive as context length grows. QSA aggregates the sequence into micro-blocks, estimates their importance at the block level, and selects only the most relevant regions. This cuts attention, compute, and indexing overhead within each layer, making the design well-suited to architectures alternating between GDN and QSA layers.
Alibaba’s published benchmarks suggest that QSA can improve the efficiency of 1M-token workloads. Compared with full attention, its attention kernel delivered speedups of up to 7.6x during prefill and 4.9x during decoding. In a cache-heavy online serving test at a 1M-token context length and with a 90% prefix-cache hit rate, Qwen3.8-Flash-Next achieved 8.6x the prefill throughput of Qwen3.7-Plus.
Running Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 #
The GB300 NVL72 features a rack-scale architecture that integrates 72 NVIDIA Blackwell Ultra GPUs into a single platform. Its large, 72-GPU NVIDIA NVLink domain enables efficient all-to-all communication at 130 TB/s, eliminating bottlenecks that appear when expert traffic must cross traditional off-the-shelf networks. Running on NVIDIA GB300 NVL72 delivers over** 16K tokens per second per GPU and over 200 tokens per second per user, **enabling developers to experiment with agentic coding applications at high throughput and low latency.
Beyond rack-scale deployment, Qwen3.8-Flash-Next also runs on local NVIDIA hardware, including NVIDIA DGX Station, NVIDIA DGX Spark clusters, and workstations with four NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition GPUs. Developers can prototype and evaluate agentic coding workflows on local hardware and scale the same model to GB300 NVL72 for production serving.
Post-train Qwen3.8-Flash-Next and serve it with your preferred inference engine #
Developers can fine-tune the model for domain-specific use cases using NVIDIA NeMo AutoModel, a PyTorch-native fine-tuning library with Day-0 Hugging Face checkpoint support. Train directly on existing checkpoints without model conversion, with support for full SFT or memory-efficient LoRA fine-tuning. Users can go a step to perform reinforcement learning using NVIDIA NeMo RL recipes.
NVIDIA supports multiple inference stacks to meet a variety of developer needs. SGLang, vLLM, and TokenSpeed provide open-source inference recipes for developers requiring greater control over performance on the NVIDIA-accelerated platform.
## Get started with Qwen3.8-Flash-Next
Try out the model from [QwenCloud.](https://www.qwencloud.com/models/qwen3.8-max)
Download the model weights from [Hugging Face](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) or [ModelScope.](https://modelscope.cn/models/Qwen/Qwen3.8-Flash-Next)