The card compares it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at their highest thinking effort, using the same harness for each. Nanbeige reports 80.9 on Terminal-Bench 2.1 (Qwen3.8-27B: 73.0) and 65.5 on DeepSWE (Qwen: 42.5). It loses on SWE-Bench Pro, 60.3 to 61.7, and Gemma4-31B leads on GPQA Diamond, 91.3 to 89.5.
Two template switches change how it behaves. enable_thinking is on by default. preserve_thinking keeps reasoning from earlier turns; the lab recommends it on for multi-turn tool use and code-agent work, and off for plain chat. For tool calls it recommends tool_call_format="xml", with JSON also supported.
Suggested settings for agent and tool-use tasks are temperature 1.0, top_p 0.95 and 65,536 max new tokens. The sample SGLang and vLLM commands use tensor parallel size 4 with the qwen3 reasoning parser and the qwen3_coder tool parser.