Technology releaseNovember 21, 2026Редакция 105.kz

Itô Labs trained Theia from scratch: what it means for LLMs inside the enterprise perimeter

Itô Labs said it trained the Theia language model from scratch: 1.22 billion parameters, its own tokenizer for Kazakh and a protective layer with role-based permissions. We look at what of this applies at an industrial enterprise and what is still a hypothesis.

Oil & GasEnergyUranium & MiningManufacturingPrivate LLM

Itô Labs trained Theia from scratch: what it means for LLMs inside the enterprise perimeter

Key takeaway. Itô Labs trained the compact Theia Eos model from scratch — 1.22 billion parameters with its own tokenizer: 1.6 tokens per Kazakh word versus 3.4–4.9 for foreign models. For a Kazakhstani enterprise this is an argument in favour of local language models, but the source contains no industrial cases and no data on on-premise delivery — the starting point should be a pilot on documentation and a check of how the model can be deployed.

What happened

Kazakhstani Itô Labs said it trained the Theia language model from scratch, without fine-tuning ready-made foreign models. According to the company, the team built the architecture, the vocabulary and the training itself; the technical details were published in a separate article, and Digital Business KZ received the “Theia Eos 1.0 Model Passport” describing the architecture, vocabulary, training stages and licences. The first model in the family is Theia Eos: 1.22 billion parameters, trained on 70 billion tokens of cleaned data in Kazakh, Russian and English, plus code and mathematics. The model is already embedded in the Opkit platform and answers user requests: creates tasks, prepares reports and drafts commercial proposals.

The developers give three reasons for training from scratch: legal independence (the weights are 100% owned by Itô Labs — company statement), its own tokenizer, and a hybrid Clocked HRM architecture where a fast part executes tasks and a separate memory module holds long-term context. According to Itô Labs, a Kazakh word takes 1.6 tokens in Theia Eos and a Russian word 1.59, against 3.4 to 4.9 for foreign models. Before launch the model was run through 48,000 work scenarios: the required action was selected in 91–95% of cases, and in the rest the model asked a clarifying question. A protective layer is claimed between the human and the model: role-based permission checks, mandatory confirmation before changing data, masking of personal data in logs, and refusal to execute commands hidden inside documents. Registration for the closed Opkit beta opened from 20 to 27 November, testing runs from 28 November to 28 December, and release is planned for 8 January 2027. Heavier Theia Selene and Theia Helio models, able to work with charts, drawings and photographs, are announced for 2027.

What this means for enterprises in Kazakhstan

An important caveat: the source contains no production case at all. All the scenarios described are office ones: tasks, reports, contracts, HR documents, exports. For an industrial enterprise this is news about a class of technology (a compact language model with local languages and role-based access), not about a ready tool for the shop floor. Applicability to production documentation is a hypothesis (assessment).

Nevertheless, the class of tasks solved here does exist at enterprises and does cost money — it is work with text, not with the process.

Sector Class of tasks for a language model What data is needed What to check in the pilot
Oil and gas Search across regulations and equipment passports, drafts of walkdown reports, parsing of defect lists Regulation texts, acts, work permits, pressure-maintenance (ПСВ) reports, maintenance (ТОиР) history Whether the answer cites the clause of the document; whether it substitutes for the engineer
Energy Work with dispatch logs, switching sheets, repair requests Logs, standard forms, requests, single-line diagrams (as text) Correctness on legacy forms and abbreviations
Uranium and mining and metals Occupational safety instructions, planned preventive maintenance (ППР), work permits, HSE reporting Instruction texts, inspection acts, orders Quality on Kazakh and Russian within one document
Plants Specifications, technical availability factor (КТГ) and repair interval (МРП) in their text part, maintenance correspondence Item master data, passports, requests, correspondence Handling of tables and long lists

A separate note on languages: Kazakh and Russian in one work permit or instruction is standard practice, and this is exactly where the tokenizer argument stops being marketing. But the claimed 1.6 tokens per word measures processing speed and cost, not the quality of an answer against a regulation.

Regulatory context: for national companies and subsoil users the key questions are personal data localisation, information security requirements and placing the model inside the perimeter. The source describes a platform with a protective layer and data masking, but does not state how the model is delivered to an enterprise (on-premise, private cloud, external API). Without that answer, any discussion of deployment remains theoretical.

How to apply: an 8–12 week pilot

  1. Data and boundary audit. Compile a list of text corpora (regulations, passports, acts, logs), determine what has no right to leave the perimeter, and define the list of roles with different access rights. Without this the pilot is a demonstration, not an acceptance test.
  2. One class of tasks with a measurable output. Not “implement AI”, but, for example, reducing the time to find the clause of a regulation currently in force, or to prepare a standard report. Measure the baseline time manually before the start.
  3. Check deployment requirements. Delivery method, licence for the weights, request logging, access separation, mandatory human confirmation before any change of data. The general approach to such systems is described on the Corporate LLMs page.
  4. A limited group with mandatory human review. 10–30 engineers, every answer used only as a draft; record the share of answers with a correct reference to the document and the share of clarifying questions.
  5. Acceptance against metrics agreed in advance. Time to prepare a document, error rate, share of outputs beyond the role's rights, cost of processing one request. Sector specifics of maintenance and dispatching are in the Industrial enterprises section. No guaranteed economic effect is promised in advance: it is measured, not assumed.

Limitations and risks

  • Model size. 1.22 billion parameters is the compact class. The source contains no public measurements on long Russian-language and Kazakh-language regulations, tables and drawings (assessment).
  • The nature of the 91–95% metric. This is the share of cases where the right action was selected across 48,000 internal platform scenarios (creating tasks, HR changes, exporting reports), not the accuracy of an answer to a question about a normative document, and not an independent evaluation.
  • Delivery and information security. The deployment method, licensing of the weights, infrastructure requirements and compliance with the information security requirements of national companies are not disclosed in the source. For an enterprise this is the first question to the vendor, not a detail.
  • Claimed rather than measured. The protective layer (role permissions, confirmation, data masking) is described by the company; no independent audit is cited.
  • Plans rather than facts. Selene and Helio with work on charts and drawings are plans for 2027; they cannot be relied on in a project decision.
  • Boundaries of application. A language model is not a means of controlling a process and does not replace the DCS (АСУ ТП); responsibility for the decision stays with the human, while vibration, SCADA trends and RUL are separate classes of tasks with different data.
  • Few facts about total cost of ownership and about how the model behaves on specific industry terminology (assessment).

Sources

  1. Digital Business KZ: Независимая отечественная ИИ-модель — разработчики рассказали, как обучили Theia с нуля · 2026-11-20T15:30:45.000Z
    Мы выбрали обучение с нуля по трем причинам. Во-первых, это юридическая независимость: модель и ее веса на 100% принадлежат Itô Labs, а для B2B и госсектора это критично.
  2. Фото: Diliff — US Capitol dome Jan 2006.jpg (CC BY 2.5, Wikimedia Commons)

This material was prepared with AI from the sources listed above. Spotted an inaccuracy? Write to info@105.kz.

FAQ

Frequently asked questions

Short, straight answers to the questions we hear most often.

Can Theia be installed inside the enterprise perimeter, on-premise?

The source does not say: it refers to the Opkit platform and its closed beta. The delivery method, licensing of the weights and infrastructure requirements must be clarified with the vendor before the pilot starts — for national companies this is a decisive factor.

Does such a model replace predictive maintenance?

No. The model works with text and with actions inside the platform: tasks, reports, document drafts. Vibration, SCADA data and remaining-useful-life estimation are a different class of tasks, and the source contains no production cases.

Why does an own tokenizer matter for a Kazakhstani enterprise?

It affects processing speed and cost: 1.6 tokens per Kazakh word versus 3.4–4.9 for foreign models, meaning more text fits into a single request. But that is a processing characteristic, not confirmed answer quality against a regulation.

Is 91–95% the model's accuracy?

It is the share of cases where, across 48,000 internal platform scenarios, the model selected the right action; in the rest it asked a clarifying question. This figure does not relate to the accuracy of answers to questions about normative documentation.

Next step

Discuss a pilot project

Tell us about your task — we'll come back with a data audit plan and an effect estimate within two business days.

  1. 1A 30-minute call: the task, your current systems, who makes the decision
  2. 2A review of data and processes, a staged plan with time and cost ranges
  3. 3A 2–4 week prototype or an 8–12 week pilot with success criteria agreed upfront

NDA from the first contact

Reply within one business day