Alibaba’s Qwen3.8‑Max Challenges GPT‑5.6 Sol and Claude Fable 5 in Benchmarks

Alibaba announced the release of Qwen3.8‑Max, the largest model in the Qwen family, featuring 2.4 trillion total parameters but activating only 95 billion parameters per inference request. The multimodal model handles text, images, and video, supports up to 1 million tokens of context, and is now available on the QwenCloud platform, with weights slated for open‑source release on Hugging Face and ModelScope next week.

The model’s programming capability is highlighted by a claim that it can start from an empty folder and complete a real‑world software project over ten days, guided by an internal “human‑in‑the‑loop” controller. Reinforcement‑learning extensions enable the model to compress a week‑long legal review into one hour and to generate complex dynamic workflows or end‑to‑end automation strategies from a single prompt. In an internal test, Qwen3.8‑Max finished an internal software‑engineering project in 16 days.

In a competitive context, the article notes that Moonshot AI’s 2.8 trillion‑parameter Kimi K3 and DeepSeek‑V4‑Flash also aim at parity with leading U.S. models. Benchmark results show Qwen3.8‑Max surpassing both Fable 5 and GPT‑5.6 Sol in all seven coding/AI/general evaluations and leading in 36 multimodal/vision tests. The full benchmark tables are shown in the images below.

API pricing is disclosed for global developers: in China, input tokens cost ¥12 per million and output tokens ¥36, with an implicit cache hit price of ¥1.5; overseas, the rates are $2 per million input tokens and $6 per million output tokens, with a $0.25 cache hit price. The article suggests that this growing pressure from Chinese providers may force OpenAI and Anthropic to lower prices and offer more flexible deployment options.