Alibaba has released the model weights for Qwen3.8-Flash-Next as a developer preview ahead of the upcoming Qwen4 family. The model is multimodal and uses a mixture-of-experts (MoE) design: the main model contains 125 billion parameters, supplemented by an additional 51 billion parameters of N-gram embeddings. Up to 6 billion parameters can be activated per token. The model has a native 262,144-token context window and can be extended to 1 million tokens using YaRN.
Architecture and scaling for long contexts
Qwen3.8-Flash-Next is intended for high-volume, context-intensive applications such as agentic coding, document processing, and tool-driven workflows. As context length increases, attention computation and KV cache memory become bottlenecks. The model addresses these with a hybrid architecture that combines two components:
- Gated DeltaNet (GDN): used in three out of four layers to continuously compress historical context into a fixed-size recurrent state, eliminating KV cache growth as sequences lengthen.
- Qwen Sparse Attention (QSA): used in the remaining layer to enable precise retrieval across the full context.
Previous sparse-attention methods relied on token-level indexers that grew more costly with longer contexts. QSA aggregates the sequence into micro-blocks, estimates block-level importance, and selects only the most relevant regions. That reduces attention, compute, and indexing overhead per layer and fits naturally with an architecture that alternates GDN and QSA layers.
Alibaba’s published benchmarks indicate QSA can improve efficiency on 1M-token workloads. Compared with full attention, the attention kernel achieved up to 7.6× speedup during prefill and 4.9× during decoding. In a cache-heavy online serving test at a 1M-token context length with a 90% prefix-cache hit rate, Qwen3.8-Flash-Next reached 8.6× the prefill throughput of Qwen3.7-Plus.
NVIDIA support and running on GB300 NVL72
NVIDIA provides best-effort Day 0 functional support through SGLang, vLLM, and NVIDIA TensorRT LLM, and validated inference runs on the NVIDIA GB300 NVL72. NVIDIA also published post-training recipes via NVIDIA NeMo AutoModel and NVIDIA NeMo RL.
The GB300 NVL72 is a rack-scale platform integrating 72 NVIDIA Blackwell Ultra GPUs into a single system. Its 72-GPU NVIDIA NVLink domain enables efficient all-to-all communication at 130 TB/s, which reduces bottlenecks that arise when expert traffic traverses conventional networks. Running Qwen3.8-Flash-Next on GB300 NVL72 delivers over 16K tokens per second per GPU at peak throughput and more than 200 tokens per second per user, enabling developers to experiment with agentic coding workloads at high throughput and low latency.
Local development and scaling path
Qwen3.8-Flash-Next 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 Max-Q Workstation Edition GPUs. Developers can prototype and evaluate agentic coding workflows locally, then scale the same model to GB300 NVL72 for production serving.
Post-train fine-tuning, inference stacks and availability
Developers can fine-tune the model for domain-specific tasks using NVIDIA NeMo AutoModel, a PyTorch-native fine-tuning library that offers Day-0 Hugging Face checkpoint support. It allows training directly from existing checkpoints without model conversion, supporting full SFT or memory-efficient LoRA fine-tuning. NVIDIA NeMo RL recipes are available for reinforcement-learning-based fine-tuning.
NVIDIA supports multiple inference stacks: SGLang, vLLM, and TokenSpeed provide open-source inference recipes for developers who need more control over performance on NVIDIA-accelerated platforms.
How to get started
The Qwen3.8-Flash-Next model can be tried on QwenCloud. Model weights are available for download from Hugging Face or ModelScope so developers can test and fine-tune the model locally or on their own infrastructure.
