Trusted models
Audited every month.
Versioned the way you read time.
Three carefully curated tiers (Fast, Pro, and Heavy) reviewed and re-selected every month by the Blankline audit team. You always get a vetted base, never whatever happened to be popular that week.
Dropstone Fast
v1.5Latency-first. Default for interactive completions, inline edits, and the moments when the loop matters more than the answer.
Dropstone Pro
v1.5The everyday driver. Strong on multi-file plans, refactors, and code that has to compile on the first try.
Dropstone Heavy
v1.5Long-context and hard cases. The tier we reach for when the spec is ambiguous or the codebase is large.
How versions work
Tier versions are written as v1.5, v1.6, and so on. The first number is the year. The number after the dot is the month.
v1.5 means the fifth audit cycle of year one. v1.6 is the next month's audit. If a tier's underlying model changes between cycles, whether promoted to a stronger candidate, swapped after a regression, or upgraded for a major capability gain, the version bumps. If nothing meaningfully improved, the version stays where it is. A version number you can trust to mean something.
Concretely: Dropstone Pro v1.5 (audited May 2026) is a different artifact from Pro v1.6 (audited June 2026) only if June's review promoted a new candidate. If the May tier still wins the June review, the version remains v1.5. You will never see a quiet swap behind the same label.
What we audit
Capability
Internal benchmark suite covering code generation, repo edits, tool use, long-context retrieval, and multi-step planning. A candidate must beat or match the incumbent at the same cost ceiling.
Safety
Refusal behavior, jailbreak resistance, prompt-injection robustness on tool-using agents, and adherence to system instructions across long conversations.
Provenance
Who trained it, who hosts it, where the inference runs, and whether the operator retains data. Anything that fails the security posture is rejected before evaluation.
Frontier models
Frontier-lab models are sometimes included in a tier when they win on a specific axis. When they are, it shows on the public model page at blankline.org/products/dropstone. The page is the source of truth for which model is behind which tier in the current cycle.
Why this matters
Running an open-source model on your own infrastructure is technically possible, but the cost is paid in three places that rarely show up on a sticker: model selection, security posture, and inference economics. Dropstone takes all three off the table.
The platform ships with three carefully curated tiers (Fast, Pro, and Heavy) that are reviewed and re-selected every month by Blankline's audit team. Every candidate model is put through a structured evaluation for capability, safety, and provenance before it is promoted to a tier, and any model whose quality or operator posture changes is replaced in the following cycle. You always get a vetted base, never whatever happened to be popular that week.
Security is treated as a hard constraint, not a feature. No prompts, code, or outputs are retained, shared with the model's original operator, or routed through inference paths controlled by Chinese or US-based labs that train on customer data. The platform runs on infrastructure that Blankline operates and audits directly, with privacy enforced regardless of which tier you use.
The price advantage is real engineering, not a subsidy. Blankline Research has rewritten the serving architecture around batching, scheduling, and memory layout to a degree most teams cannot justify building in-house, which is why a Dropstone Pro seat costs a fraction of what equivalent capability costs at a frontier lab.
In effect, you get the model selection of a research group, the security posture of an enterprise platform, and the unit economics of a system engineered from the ground up to be cheap, without having to build any of it.