- Add Terminal-Bench 2.1 and DeepSWE scores to model card (775335f55fdf788f4e66637f210f7a9085ea560f)
- Align Terminal-Bench denominator to full 89-task set (d2df1d2e5d683a864b7fa6447dde61452a8eab4c)
- Make Evaluation section consistent across benchmarks (b4057c06e3c8cf118a17345114022f233b542bed)
chat template: early break in tool result reordering check
Following the optimization in [this SGLang PR](https://github.com/sgl-project/sglang/pull/36957), added four `break` statements so the loop stops early once it detects that the results can't be reordered.
