The main evaluation uses five instruction-tuned open-weight models from five tokenizer lineages; only Ministral-3-8B-Instruct-2512 and Qwen3.8-27B define a tool role, and Qwen is the only listed model with template-emitted control identifiers left unmarked as special.

model: gemma-4-31B-it · owner: google

model: gemma-4-31B-it · toolRole: no

model: Llama-3.1-8B-Instruct · owner: meta-llama

model: Llama-3.1-8B-Instruct · toolRole: no

model: Ministral-3-8B-Instruct-2512 · owner: mistralai

model: Ministral-3-8B-Instruct-2512 · toolRole: yes

model: Qwen3.8-27B · owner: Qwen

model: Qwen3.8-27B · toolRole: yes

model: gpt-oss-20b · owner: openai

model: gpt-oss-20b · toolRole: no