Read the date and the status before the number
The public research snapshot is dated 11 October 2026 and labels its figures as collected data rather than model results. Research and design remain in progress; the first coding pilot and evaluation are planned.[1]
This is Kooya’s explanation of its public roadmap, not an independently audited inventory or financial review. A target, a measured quantity and a result answer different questions.
Parameters and tokens measure different things
A parameter is a numerical value within a model. A token is a unit of data processed by its tokenizer; it is not necessarily one word. A count of billions of parameters therefore describes a different thing from billions of tokens.
Hoffmann and colleagues’ 2022 research studies the relationship between model size, training tokens and a compute budget. It supports examining those dimensions together rather than treating parameter count alone as sufficient. Its experiments do not establish a recipe or an outcome for Pino SI’s planned adaptation.[2]
For this roadmap, the simple reading rule is to keep the units visible. A data inventory cannot be converted into model size, and a model-size target cannot tell a reader how much material has actually been used in a completed training run.
Understand the core model-size targets
The regional-languages and knowledge focus lists a recommended first-version target of 30–35 billion parameters and a future target of 70 billion or more. STEM, coding and defensive cybersecurity each list 14–32 billion for the first version and 70 billion or more as the future target. Formal math has a separate 7–14 billion first-version target and a 32 billion future target.[1]
These are planning ranges. They are not a catalogue of released checkpoints or a claim that the largest figure has been achieved. The “future” label is material: removing it would change the meaning of the public disclosure.
Preserve the labels on the data counts
The site reports 3.16 billion measured language tokens, including 2.37 billion Tagalog tokens. It separately shows approximately 50 billion math tokens; an estimated 8–12 billion science tokens; an estimated 3–5 billion formal-math tokens; an estimated 40–70 billion code tokens; and an estimated 4–7 billion tokens in the defensive scope.[1]
The Tagalog figure is included in the language total, so adding it again would double-count that portion. The other quantities have different status labels and should not be combined into an apparently precise grand total without a documented basis.
This note publishes aggregate figures already on the site. It does not identify training collections or their sources. More importantly, an inventory count does not say that all material is approved, processed or used in a completed training run.
Read small pilots as a different stage of work
Agriculture, supply chain, biosafety and preparedness, and critical infrastructure are labelled planned proof-of-concept areas. Each lists 4–8 billion parameters as a first-version target and 14–32 billion as a future target. Agriculture shows approximately 0.3 billion tokens plus structured data. Supply chain has no token total. Biosafety shows an estimated 0.1 billion tokens, marked pending. Critical infrastructure lists tools and simulators rather than a text corpus.[1]
These categories cannot sensibly be forced into one text-volume comparison. A simulator and a token inventory describe different research resources. The public plan calls for a focused task and a checkable result before expansion.
For readers, the useful distinction is scope: these are exploratory directions, not evidence of operating farming, health, logistics or infrastructure services.
The small-model pilot note turns these four planned domains into examples of specific questions a future experiment could test.
Distinguish fundraising from money already invested
The site identifies US$5 million as its fundraising target and reports US$50,000 invested by Kooya from its own funds. The two numbers are presented in different roles. The target is not described as funding already raised.[1]
A reader should therefore avoid converting the target into an available research budget or an announced valuation. The reported own investment is also not, by itself, an independently audited expenditure breakdown. This article adds no claim about investors, commitments or allocation of future funds.
Our view is that research planning benefits from keeping resources and milestones linked, while preserving their uncertainty. The public figures communicate an ambition and a starting contribution. They do not determine a model’s capability or guarantee the completion of every roadmap item.
Ask what evidence would support the next statement
The next useful question is not which number looks largest, but what it demonstrates. Inventory figures describe data progress. A trained model would need its own documentation. A capability statement would need evaluation conditions and results. Deployment would require another account of intended users and operating limits.
The present roadmap is a dated research snapshot. Its value lies in showing the current direction and the labels attached to it. Readers can take the ambition seriously while retaining the distinction between planned scale, reported resources and evidence that still has to be produced.