Qwen3.8-Flash-Next Release

Alibaba Qwen released Qwen3.8-Flash and open-sourced Qwen3.8-Flash-Next, an early preview of the Qwen4 architecture intended for inference framework adaption before the full family launch. - Efficiency vs. Qwen3.7-Plus (397B/A17B): 1/3 active parameters, 1/3 training tokens, approximately 1/9 training FLOPs; leads on 8 of 14 pretraining benchmarks, all other scores within 2.6 points. - Training stability: stress test at 4× optimal learning rate shows the old recipe (AdamW + Qwen3.5) produces frequent loss spikes, while the new recipe (Muon optimizer + gated residuals) has zero spikes; production training never required qk-clip. - Post-training official benchmarks: DeepSWE 1.1 58.7 (Qwen3.7-Plus 16.5, DeepSeek-V4-Flash 54.4); SWE-bench Pro 62.5; SWE-bench multilingual 81.0; CoWorkBench 73.9, JobBench 55.7, Toolathlon 73.5; GPQA Diamond 91.7, LiveCodeBench v6 91.9, IFBench 81.3; AndroidWorld 84.5, MathVision with code interpreter 95.7, OSWorld 2.0 partial 52.3, LVBench 76.6. - NVIDIA provided Day 0 support with NeMo AutoModel and NeMo RL fine-tuning recipes, plus integration with SGLang, vLLM, and Lightseekorg TokenSpeed.
Key insights
Early Preview of Qwen4 Architecture Aims for Ecosystem Readiness
Qwen3.8-Flash-Next is explicitly labeled as an early preview of the Qwen4 architecture, open-sourced ahead of the full family to allow inference frameworks like vLLM and SGLang, and infrastructure developers, to complete adaptation before the official Qwen4 release. NVIDIA confirmed Day 0 support with fine-tuning recipes and integration.
Breakthrough Efficiency: Lower Params Yet Strong Competitiveness
Compared to its predecessor Qwen3.7-Plus, Flash-Next uses only 1/3 the active parameters, 1/3 the training tokens, and approximately 1/9 the training FLOPs, yet leads on 8 of 14 pretraining benchmarks and trails by at most 2.6 points on the rest. This drastic reduction in compute while retaining competitive performance highlights architectural improvements.
Muon Optimizer and Gated Residuals Eliminate Training Instability
A reproducible stress test applying 4× the optimal learning rate demonstrated that the old AdamW + Qwen3.5 recipe suffers frequent loss spikes, whereas the new Muon optimizer with gated residuals exhibits zero spikes. Furthermore, the production training pipeline avoided any explicit qk-clip, indicating a more stable and robust training method.
Agentic Coding and Long-Horizon Benchmarks Show Clear Leadership
On agentic coding tasks, Qwen3.8-Flash-Next scored DeepSWE 1.1 58.7, far above Qwen3.7-Plus’s 16.5 and higher than DeepSeek-V4-Flash 54.4. It also achieved SWE-bench Pro 62.5, multilingual 81.0, and strong long-office/agent results such as CoWorkBench 73.9, JobBench 55.7, and Toolathlon 73.5—all notably ahead of compared models.
Multimodal Capabilities Extend to Visual and Video Tasks
The model reached AndroidWorld 84.5, MathVision with code interpreter 95.7, OSWorld 2.0 partial 52.3, and long-video LVBench 76.6, demonstrating strong performance across device operation, mathematical vision, OS interaction, and video understanding.