Sber has opened registration for the international online artificial intelligence competition AI Journey Contest 2026. This year's prize pool exceeds 10 million rubles, and the submission period will run until October 20, 2026, inclusive. Participants will need to choose one of five practical tasks on the GitVerse AI Cup platform — ranging from detecting hallucinations in LLM responses to diagnosing coordination in multi-agent swarms. Participation is free, either individually or in a team of 2–4 people.

!image:format(jpg)/YXJ0aWNsZXMvaW1hZ2UvMjAyNi84L2xvcmktMDA1MDUxOTI2NC1iaWd3d3dfTUZ6MDJJai5qcGc.jpg)

image

What happened

The competition will be held in an online format, and the tasks for it were prepared by the Sber AI team together with the competition's partner, cloud provider Cloud.ru. The Guardian of Truth track is dedicated to detecting hallucinations in the responses of large language models. Green Challenge requires building a control model for a humanoid robot within the Physical AI and VLA direction. In the 3D DIT Distillation task, participants need to speed up the 3D generator Kandinsky 3D through distillation of a diffusion model. Vision-Action Data Factory involves generating training data for fine-tuning the robotic model SmolVLA, while Coordination Referee — creating an arbiter model that diagnoses the quality of coordination in multi-agent swarms. For each track on GitVerse AI Cup, public datasets and leaderboards are open, participants must be over 18 years old, and the award ceremony for the winners will take place at the AI Journey conference in Moscow.

Context

The choice of tracks itself serves as a signal: the organizer publicly identifies five directions it considers the frontier of agentic and robotic systems. Two of the five tasks are related to VLA models and Physical AI, with one entirely dedicated to data generation for SmolVLA — this indirectly indicates that the organizers see the bottleneck in robotic models as training data, not just architecture. Guardian of Truth and Coordination Referee relate to evaluation methodology: a hallucination detector and a coordination arbiter are themselves models and can therefore become a source of errors, so the outcome of such tasks depends on the design of reference labels and held-out test sets. The 3D DIT Distillation track lies in the area of inference engineering: distillation of a diffusion model traditionally means a trade-off between speed and generation quality.

Why this matters for the industry

For the industry, the contest is valuable primarily as a public benchmark: the datasets and leaderboards of GitVerse AI Cup provide a comparable external check for hallucination detectors, multi-agent arbiters, accelerated 3D generators, and synthetic data pipelines. The five tracks effectively describe in-demand product components — a hallucination filter for LLM responses, a coordination arbiter for agentic systems, a fast version of Kandinsky 3D, and a training data generator for robotics. The involvement of Cloud.ru indicates the cloud infrastructure available for participants' solutions. For startups and product teams, the list of tracks suggests where unresolved pain points are concentrated, although the competition solutions will remain separate features rather than ready-made products. After the submission deadline, the accumulated leaderboards will allow judging the real complexity of the tasks, and strong approaches, if the organizers publish them, could become public baselines for subsequent development.

Why this matters for users

For participants, this is a chance to work on industrial tasks and get a ready-made pet project for their portfolio: for example, a prototype of a hallucination detector on Guardian of Truth or training a robot in a simulator on Green Challenge. The public datasets of the tracks can be used outside the competition as well — as an external evaluation loop for their own models, including hallucination detectors, multi-agent arbiters, and data generation pipelines. The 2–4 person team format sets a compact sprint scope, while individual participation allows moving at one's own pace. The window from registration to the deadline is less than a month, so it is worth choosing a track and checking access to datasets and the leaderboard immediately.

What is still unknown / limitations

The target metrics for 3D DIT Distillation are not disclosed in the open materials: neither the target speed nor the requirements for maintaining quality after distillation are known, and without them, a correct comparison of solutions remains an open question. For Guardian of Truth and Coordination Referee, it is unclear how the reference labels and held-out test sets are structured, and this determines whether circularity of self-evaluation can be avoided. It is too early to speak of a reproducible industry benchmark until the methodology and datasets are published and it is confirmed whether strong solutions will be made public as baselines. Finally, reading the list of tracks as a map of Sber AI's priorities is an interpretation based on indirect signs, not an official statement by the organizers.

Sources

Author

Look at AI, editorial team