Alibaba to Release Next-Gen “Qwen3.8-Flash-Next” Model for Qwen4 on August 27 — BigGo Finance

Alibaba’s AI model development team Qwen will release “Qwen3.8-Flash-Next,” which adopts a new architecture designed for the next-generation “Qwen4” model, as an open-weight model on August 27 at 12:00 AM Japan time. The goal is to give developers and users early access to design improvements so they can prepare for the new architecture ahead of Qwen4’s official release.

Qwen3.8-Flash-Next is a multimodal MoE (Mixture-of-Experts) model built on top of the Qwen4 architecture. MoE is a mechanism that divides the model’s internal structure into numerous small networks with different roles, activating only the necessary parts for each input to reduce computational load.

Details such as parameter count and licensing have not been disclosed at the time of writing. However, there is precedent for the “Flash” designation. The cloud-based “Qwen3.5-Flash” was positioned as an enhanced version of the open-weight “Qwen3.5-35B-A3B” (35 billion total parameters, 3 billion active parameters). Whether this Qwen3.8-Flash-Next will be equivalent to a similarly scaled “Qwen3.8-35B-A3B” remains unclear at this point.

The Qwen team announced the release on its official X (formerly Twitter) account with the post: “The next Qwen wave is coming.”

Regarding the Qwen3.8 series, overseas tech media outlets have begun publishing comparison articles on performance and cost. According to those reports, the open-weight version of Qwen3.8 reportedly features 2.4 trillion total parameters with 95 billion always-active parameters, priced at $2 per million input tokens and $6 per million output tokens. Processing speed is reported at 47–48 tokens per second — somewhat slower than competing closed-weight models — but it is evaluated as well-suited for large-scale codebase processing, long-term agent-based tasks, and cost-sensitive projects.

However, it should be noted that these overseas media reports pertain to the Qwen3.8 series as a whole and may not necessarily match the specifications of the Qwen3.8-Flash-Next being released this time.

Qwen3.8-Flash-Next Release Information

Note: Qwen3.5-Flash was an enhanced version of Qwen3.5-35B-A3B with 35 billion total parameters and 3 billion active parameters, but this release is not necessarily of the same scale.

Qwen is a family of large language models developed by China’s Alibaba Group, which has actively pursued open-weight releases. Since 2025, the series has expanded in performance and functionality across generations — Qwen2.5, Qwen3, and Qwen3.5 — and this Qwen3.8-Flash-Next serves as a bridge to Qwen4 along that trajectory.

The early open-weight release is seen as a strategy to accelerate ecosystem formation. By allowing developers to familiarize themselves with the new architecture and provide feedback, the team can improve the quality of the final Qwen4 release. On the other hand, with parameter count and licensing terms still undisclosed, companies considering commercial use currently have limited information on which to base decisions.

Competition between open-weight and closed-weight models is intensifying in the AI model market. Qwen’s open strategy, alongside Meta’s Llama series and Mistral, aims to capture the developer segment that prioritizes customizability and cost efficiency. The release of Qwen3.8-Flash-Next serves as a strategic move to build anticipation for Qwen4 within this competitive landscape.