- Published on
Experiment with Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 for Agentic Coding
- Authors
- Name
- DeepHorus
Alibaba has released the model weights for its new Qwen3.8-Flash-Next AI model as a preview of the upcoming Qwen4 architecture. This experimental multimodal model is designed specifically for agentic coding and other context-intensive tasks. It features a 125-billion-parameter main model based on a mixture-of-experts (MoE) architecture, activated with 6 billion parameters per token. The release is aimed at developers for early experimentation and evaluation ahead of the full Qwen4 launch.
The model's design addresses a core challenge in long-context AI applications: the computational and memory bottlenecks of growing attention mechanisms. It employs a hybrid architecture combining two innovative techniques. Gated DeltaNet (GDN) layers compress historical context into a fixed-size state, eliminating KV cache growth. Qwen Sparse Attention (QSA) layers then enable efficient retrieval across the full context window.
This architecture enables a native context window of 262,144 tokens, which is extensible to 1 million tokens using YaRN technology. Published benchmarks show the QSA system can deliver up to 7.6x speedups during prefill and 4.9x during decoding compared to full attention. In online serving tests at a 1-million-token length, it achieved 8.6x the prefill throughput of the previous Qwen3.7-Plus model.
The model is validated for high-performance inference on NVIDIA's latest GB300 NVL72 platform. This system integrates 72 Blackwell Ultra GPUs into a single rack with massive 130 TB/s NVLink bandwidth. The combination delivers over 16,000 tokens per second per GPU and over 200 tokens per second per user. This performance is critical for real-time agentic coding workflows requiring high throughput and low latency.
Developers can now access and fine-tune the model weights from Hugging Face or ModelScope. NVIDIA is providing full support through tools like NeMo AutoModel for fine-tuning and NeMo RL for reinforcement learning. Multiple inference stacks, including SGLang and vLLM, are also supported for flexible deployment.
The release signals a strategic push in the race for more capable, context-aware AI models. By making this advanced architecture available early, Alibaba and NVIDIA are empowering developers to build and test the next generation of AI-powered software development tools. The focus on agentic coding points toward a future where AI assistants can handle extensive codebases and complex, multi-step tasks with greater efficiency.
