Alibaba has introduced its latest flagship, Qwen3.7-Max, designed for long-running agent tasks. During the demonstration, the model autonomously performed 1,158 tool calls in 34.7 hours, with the SGLang Triton Kernel, a component of the popular open-source library for large model inference, providing a tenfold speedup. For comparison, other models like DeepSeek V4 Pro and Kimi K2.6 achieved speed increases of 3.3x and 5x, respectively. This is reported by Habr.com .
Qwen3.7-Max was trained in over 8,200 diverse environments, enabling it to plan multi-step tasks, call tools, and respond to their results. According to Alibaba, the increase in environments improved the model's average ranking across eight agent benchmarks almost linearly, rising from 9th place in the base version to 3rd in the final version.
In 12 selected public benchmarks, Qwen3.7-Max leads in almost all categories. For example, in Terminal-Bench 2.0, it scored 69.7 points compared to 65.4 points for Claude Opus 4.6 Max Thinking. However, it lags slightly in long coding tasks on the NL2Repo benchmark, where Claude outperforms it by 0.4 points. Notably, some new competitor versions were not included in the comparisons.
Qwen3.7-Max is available for free via the company's chatbot and API, priced at $2.5 per 1 million input tokens and $7.5 per 1 million output tokens. However, open weights are not being announced in line with previous Max versions.
Read “Zamin” on Telegram!
