On August 26, Alibaba rolled out and open-sourced its newest model, Qwen3.8-Flash, within the Qianwen model series. This model is built on a cutting-edge architecture, boasting a total of hundreds of billions of parameters. However, it strategically activates only 6 billion parameters (6B) to outperform Claude Opus 4.6, thereby establishing a new standard for model efficiency. Owing to extensive optimizations in both architecture and training processes, the training expenses for Qwen3.8-Flash have been substantially slashed by nearly 90% in comparison to Qwen3.7-Plus. Additionally, the inference costs have seen a significant decrease, with input costs plummeting to as low as 1 yuan per million Tokens and output costs fixed at 3 yuan, rendering it merely one-third the price of DeepSeek-V4-Flash at its most economical point. Qwen3.8-Flash is set to debut on 'Qianwen Office' on the same evening, enabling developers and businesses to tap into the new model's API services via the Qianwen AI platform.
