When open models learned to act
Tool calling took three years to become universal. Reasoning took two.
Every model that talks to a program does it through a chat template — the small piece of Jinja that turns a conversation into tokens. The template is where a model’s abilities become concrete. If it has somewhere to put a tool call, the model was trained to make one. If it has nowhere, it was not.
That makes templates a record of when open models stopped being things you talk to and started being things that do work. We read 175 of them, from Jul–Dec 2023 to Jul–Dec 2026.
Tool calling took three years
In Jul–Dec 2023, none of the releases we can assess had anywhere to put a tool call. By Jul–Dec 2026, 94% did. The first appeared in Jan–Jun 2024, in 2 of 14 releases.
The climb was not smooth. Adoption fell back in Jan–Jun 2025, to 44%. That dip is composition rather than retreat: the period carried a large ERNIE release wave and a cohort of vision-language models, and vision models came to tool calling late. In 2025 they showed it at a lower rate than text-only releases. In 2026 they overtook them.
Reasoning took two years
The sharper curve is the second one. Reasoning support does not appear at all until Jan–Jun 2025 — not rare, absent. Within 4 half-years it reaches 94%.
Tool calling had to wait for an ecosystem: a function-calling convention, parsers, frameworks willing to route a model’s output back into code. Reasoning needed none of that. It is a training decision and a few tokens in a template, so once it worked it could spread at the speed labs could retrain.
Why it matters
An agent builder choosing an open model in 2024 was choosing from a minority that could act at all. In Jul–Dec 2026 the question has inverted: tool calling is the default, and its absence is the thing worth checking for. The scarce capability moved on.
Method, and what this does not show
- Derived from 175 of 422 releases in the register. The rest publish no chat template, so they cannot be assessed — that is not evidence of missing support.
- Quantisations and conversions are excluded. They inherit their parent’s template and would double-count it.
- Release date is the model repository’s creation date, the best public proxy available. It can trail an announcement.
- The register tracks named organisations plus models above a download threshold, so this is a sample of notable releases, not of everything published.
- Detection reads the template for tool-call, reasoning and media handling. It found 5 false positives — autotrust/JEV-35B, autotrust/GEV-26B-Decide, autotrust/JEV-9B, Mapika/decider-2b, Qwen/Qwen3-Embedding-0.6B, non-generative models inheriting a generic template. At 175 assessed that is 2.9%.
- Periods with fewer than three assessable releases are omitted.
Every figure here is computed at build time from the register, which is CC BY 4.0. The chart cannot drift from the data behind it.