Alibaba’s Qwen to Open-Source Qwen3.8-Flash-Next, Previewing Qwen4 Architecture · TechNode
The open-weight preview uses 125B parameters but activates only 6B per token, while Alibaba weighs charging its largest commercial users.
- Alibaba released Flash-Next, an open-weight preview of the architecture intended for Qwen4, carrying 125B parameters while activating only 6B for each token.
- Compared to Qwen Plus, which activates 17B, the model runs on roughly a third of the active compute using a sparse attention scheme and gated residual mechanism.
- TNW noted this month the company provides no proof for its performance claims, and the model card reveals GPT-4o graded the Humanity Last Exam score rather than the benchmark's own grader.
- The weights sit on Hugging Face under a license that may not qualify for the European Act's open-source exemption, as Recital 103 suggests monetized components lack documentation duty relief.
- Thomson Reuters built its model on Qwen, inheriting license terms that Alibaba intends to monetize; European firms are not waiting for clarification on these regulatory questions.
16 Articles
16 Articles
Qwen4's architecture is here early, firing 6B parameters out of 125B
Alibaba’s Qwen team has released Qwen3.8-Flash-Next, an open-weight preview of the architecture it intends to use for Qwen4, carrying 125B parameters but activating only 6B for each token. Its licence may not qualify for the EU AI Act’s open-source exemption. Alibaba’s Qwen team has published the architecture it plans to build Qwen4 on. Qwen3.8-Flash-Next carries […] This story continues at The Next Web
Alibaba’s Qwen to open-source Qwen3.8-Flash-Next, previewing Qwen4 architecture · TechNode
Alibaba’s Qwen team is scheduled to open-source Qwen3.8-Flash-Next and an FP8 version at 11 p.m. Beijing time on Aug. 26. The model is described as a multimodal mixture-of-experts system built on the next-generation Qwen4 architecture. Qwen said the early release is intended to give the developer community an opportunity to prepare for the full Qwen4 model family. The team has not yet disclosed further details about the model’s specifications or…
Alibaba's AI research team, Qwen, released its AI model "Qwen3.8-Flash-Next" on August 26, 2026. Qwen3.8-Flash-Next was developed using the next-generation architecture planned for the Qwen4 series, successfully building a high-performance model while reducing training costs. A quantized version using Unsloth has also already been released. Read more...
Alibaba Open-Sources Qwen3.8-Flash with 6B Active Parameters at One-Ninth Training Cost
Late on August 26, Alibaba’s Qwen team released Qwen3.8-Flash, an open multimodal mixture-of-experts model that it describes as a technical preview of the next-generation Qwen4 architecture. The name still carries a 3.x tag, but the design is a new generation, released early so the community can validate the architecture ahead of the full Qwen4 family. The headline is cost. The main model has 125 billion parameters with an additional 51 billion-…
Coverage Details
Bias Distribution
- 34% of the sources lean Left, 33% of the sources are Center, 33% of the sources lean Right
Factuality
To view factuality data please Upgrade to Premium
