Exploring Alibaba’s Qwen3.8-Flash-Next: A Leap in Multimodal AI

Alibaba has recently unveiled the model weights for Qwen3.8-Flash-Next, positioning itself on the cutting-edge of AI development. This release serves as a preview of the anticipated Qwen4 architecture, offering developers an opportunity to explore and evaluate a robust multimodal mixture-of-experts (MoE) model. With a substantial 125 billion parameter main model supported by an additional 51 billion N-gram embeddings, Qwen3.8-Flash-Next operates with an impressive 6 billion parameters activated per token. One of its standout features is a native context window of 262,144 tokens, expandable to an astonishing 1 million tokens using YaRN.

NVIDIA, a key player in advanced computing, provides foundational support for Qwen3.8-Flash-Next through various tools, including SGLang, vLLM, and NVIDIA TensorRT LLM. Through validation on NVIDIA’s GB300 NVL72, users can expect optimized inference capabilities along with post-training recipes available from NVIDIA NeMo AutoModel and NeMo RL.

Architectural Innovations for Long-Context Inference

Designed for demanding applications such as agentic coding, document processing, and tool-driven workflows, Qwen3.8-Flash-Next has made significant strides to address the challenges of attention compute and memory bottlenecks that arise when processing extensive context. At the heart of this model lies a hybrid architecture that integrates Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA).

A notable feature of this architecture is its use of GDN in three out of every four layers, which continuously compresses historical context into a fixed-size recurrent state. This innovative approach effectively eliminates the growth of key-value (KV) caches, a common issue as sequence lengthens. The remaining layers incorporate QSA to enable precise retrieval across the entire context.

Traditional sparse-attention techniques often struggle with computational costs due to token-level indexers when context length increases. However, QSA employs a novel aggregation method, grouping sequences into micro-blocks and evaluating their importance at the block level. This intelligent selection of relevant regions results in a remarkable reduction in attention compute and indexing overhead, enhancing the model’s efficiency in managing large contexts.

Benchmark results published by Alibaba point towards substantial efficiency gains, especially in handling 1 million-token workloads. The attention kernel of QSA reported speed improvements of up to 7.6x during prefill and 4.9x during decoding. In scenarios with a 90% prefix-cache hit rate, Qwen3.8-Flash-Next showcased an impressive 8.6x increase in prefill throughput over Qwen3.7-Plus.

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

The NVIDIA GB300 NVL72 is a cutting-edge rack-scale architecture featuring 72 NVIDIA Blackwell Ultra GPUs, creating a robust platform for AI applications. This configuration supports efficient all-to-all communication at an extraordinary 130 TB/s, effectively eliminating bottlenecks commonly encountered when expert traffic crosses traditional network boundaries. For developers, running Qwen3.8-Flash-Next on this platform translates to over 16,000 tokens per second per GPU and more than 200 tokens per second per user. These capabilities enable high-throughput and low-latency experimentation in agentic coding applications.

Moreover, Qwen3.8-Flash-Next isn’t limited to rack-scale deployment; it can also be effectively run on local NVIDIA hardware. Systems like NVIDIA DGX Station and DGX Spark clusters, as well as workstations featuring four NVIDIA RTX PRO 6000 Blackwell Max-Q GPUs, allow developers to prototype and assess various agentic coding workflows before scaling their models to GB300 NVL72 for production.

Post-Train Qwen3.8-Flash-Next and Serve It with Your Preferred Inference Engine

To cater to specific application needs, developers can fine-tune Qwen3.8-Flash-Next for domain-specific use cases using NVIDIA NeMo AutoModel. This PyTorch-native fine-tuning library facilitates direct training on existing checkpoints without requiring a model conversion. It also supports full supervised fine-tuning (SFT) or memory-efficient approaches like LoRA fine-tuning. Furthermore, users looking to enhance model performance can explore reinforcement learning options using NVIDIA NeMo RL recipes.

To support a diverse range of development requirements, NVIDIA offers an array of inference stacks. Through SGLang, vLLM, and TokenSpeed, developers can utilize open-source recipes to optimize their models to deliver the performance they need on NVIDIA-accelerated platforms.

Get Started with Qwen3.8-Flash-Next

For developers eager to dive into this groundbreaking model, QwenCloud serves as a platform to trial the capabilities of Qwen3.8-Flash-Next. Additionally, model weights are readily available for download from Hugging Face or ModelScope, simplifying the start of your journey into advanced multimodal AI.