Qwen3.8-Flash-Next

Qwen3.8-Flash-Next is an open-weight, multimodal Mixture-of-Experts (MoE) language model released by Alibaba's Qwen team. Launched as an early architectural preview of the future Qwen4 family, the model is designed to maximize computational efficiency while delivering high performance in agentic workflows, coding, and tool-driven tasks. It employs a sparse activation strategy that activates only 6 billion main-model parameters per token out of a total parameter pool of approximately 180 billion.

Architecture and Innovation

The model introduces several structural advancements intended to reduce the cost of long-context processing. It utilizes a Hybrid Attention mechanism that combines Gated DeltaNet (GDN) for efficient history compression with Qwen Sparse Attention (QSA), which operates at a micro-block level to reduce latency during inference. Additionally, it incorporates Gated Residual (GR) streams to improve cross-layer information flow and an N-gram Embedding table of 51 billion parameters to scale model capacity without increasing the active compute requirement.

Capabilities and Context

Qwen3.8-Flash-Next supports a native context window of 262,144 tokens, which can be extended to 1,000,000 tokens using YaRN. As a multimodal system, it natively understands text, images, and video, supporting tasks ranging from document analysis to mobile-use automation. The model also features a native Thinking mode for advanced reasoning, which can be controlled via a reasoning_effort parameter to adjust the depth of internal processing before a final response is generated.

Benchmarks released by Alibaba indicate that Qwen3.8-Flash-Next shows significant improvements in software engineering tasks, reportedly outperforming contemporary models like Claude 4.6 Opus on evaluations such as SWE-bench Pro. Its training recipe utilizes the Muon and AdamW optimizers, refined specifically for the new architecture's scaling laws to ensure robust convergence at high learning rates.