MTS AI выпустила Cotype Pro 3 и Cotype Light 3: как учат модели агентским навыкам

MTS AI released the third generation of its Cotype language models in July 2026 — the flagship Cotype Pro 3 and the lightweight Cotype Light 3. Along with the release, the company revealed how it collects real task-execution traces to train the models in agentic skills — the ability to complete multi-step scenarios without failures.

What MTS AI released

MTS AI has updated the Cotype lineup to its third generation: the flagship Cotype Pro 3 and the lighter Cotype Light 3, which launched slightly ahead of the flagship. The company positions both models as a key component of corporate AI agents and multi-agent systems — that is, as the "brain" that manages chains of actions rather than simply answering single queries.

• Developer — MTS AI
• Flagship model — Cotype Pro 3
• Lightweight version — Cotype Light 3 (released slightly ahead of the flagship)
• Purpose — core of corporate AI agents and multi-agent systems
• Training focus — completing multi-step scenarios without errors

Why the data can't be made up

Traces are recorded trajectories of how a model solves a real task step by step: it calls a tool, reads the result, and chooses the next step. It's on exactly this kind of data that MTS AI trains Cotype's agentic skills, and, as the company emphasizes in the headline of its own breakdown, this data "can't be made up."

Generating synthetic data for single-turn answers is easy, but realistic multi-step scenarios are not: they depend on genuine edge cases, tool failures, and branching points that can't be artificially reproduced. That's why an agent's quality hinges not on model size but on the quality and authenticity of the collected traces. According to MTS AI, without this kind of training a model "stumbles" on long chains of actions.

Where Cotype 3 is used

Cotype Pro 3 and Cotype Light 3 are designed primarily for corporate scenarios — acting as an executor and orchestrator within multi-agent systems. In this kind of architecture, one model breaks a task down into steps, calls the necessary tools, and passes intermediate results to other agents, which makes the cost of a single error at any step high.

MTS AI calls reliability in multi-step scenarios a mandatory condition for its AI agents to function. The lightweight Cotype Light 3, meanwhile, is designed for more cost-effective deployment, while the flagship Pro 3 is built for the most complex tasks.

"...how we teach our models to carry out multi-step scenarios without stumbling, which is absolutely essential" — from MTS AI's technical blog on

Habr.

What this means

Russian LLMs are increasingly being tuned not for "chit-chat" but for work within agentic systems, where the reliability of long chains of actions matters most. MTS AI's bet on collecting real traces shows that competition is shifting from model size to the quality of data used to train agentic skills.

Frequently asked questions

What are Cotype Pro 3 and Cotype Light 3?

They are the third generation of MTS AI's language models: Cotype Pro 3 is the flagship version, Cotype Light 3 is the lightweight one. Both are built as a key component of corporate AI agents and multi-agent systems.

Why does MTS AI collect traces when training Cotype?

Traces are real, step-by-step trajectories of solving tasks. MTS AI trains agentic skills on them so the model can get through multi-step scenarios without failures; according to the company, this kind of data can't be reliably synthesized.