How Dropstone chooses its models
Every month Blankline re-evaluates the open-weight models available for each Dropstone tier against three tests, how well it does the work, what it costs to serve, and how safely it behaves with tools. A model has to beat the one already in the slot.
Fast, Pro, and Heavy are filled by a monthly evaluation rather than a permanent choice. This page covers what is measured, who does it, and what it takes to win a slot.
Who decides?
The evaluation is run inside Blankline, the company that makes Dropstone. It is a standing commitment rather than a one-off: the lineup is reconsidered every month, including the months where the answer is that nothing should change.
What is measured?
Every candidate is tested on three things.
| What is measured | What it looks at |
|---|---|
| How well it does the work | Code generation, editing an existing repository, tool use, retrieving from long documents, and planning across several steps |
| What it costs to serve | The measured cost of a typical heavy coding turn, not the headline rate of the model |
| How safely it behaves with tools | Following instructions across a long conversation, refusing what it should refuse, and holding up against prompt injection when it is allowed to act |
How a model takes a slot
A candidate has to beat or match the model already in the slot. Being newer is not a qualification on its own.
The three tiers are judged separately, because the work differs, so a model can win one slot and lose another. A model that loses a slot can also move down rather than leave: the model that held Heavy before Kimi K3 took it moved down to Pro instead of being dropped.
What happens after a pick
Once a tier's pick changes, the version moves and the new lineup ships. Model versions and the monthly lineup covers the numbering, and the release notes record what changed.
Nothing else moves with it. Your allowance, your memory, and your conversations are unaffected.
Where the reasoning is published
Blankline publishes the reasoning behind a lineup, not only the outcome. The write-ups and the benchmark material live at blankline.org/research, and the Dropstone blog covers the lineups in plainer language.
If a cycle changes nothing, there is nothing to announce, and the version stays where it is.
Related articles
- Which AI models does Dropstone use?Dropstone runs three tiers. On the 1.8 lineup, Fast runs DeepSeek V4.1 Flash, Pro runs GLM-5.3 Flash, and Heavy runs Kimi K3, re-evaluated every month.
- Model versions and the monthly lineupDropstone numbers each lineup with the year and the month it was cut, and a new version ships only when a tier's pick changes. How to read the number, and which lineup you get.
- Why Dropstone runs open-weight modelsThe models behind Dropstone Fast, Pro, and Heavy are open-weight, so the weights are published and anyone can check which model is behind a tier. Here is what that gives you, and what it does not mean.
- Which model am I using?The button at the foot of the composer names the tier, the lineup version and the effort level, and your usage page records which model handled each request.
- Choose a modelHow to pick between Dropstone Fast, Pro, and Heavy, where the selector lives in chat and in the CLI, and why the name Fast describes what a tier costs rather than how fast the model inside it runs.