SberAI releases FRIDA-Decisions, an 823M encoder decision model for Russian

Source · Model card: FRIDA-Decisions

SberAI’s УЭСМО team published FRIDA-Decisions on 2 October 2026 under the ai-forever organization. It is a decision model built on FRIDA, an 823M-parameter T5 encoder, and answers choice, score, noul and ranking questions about a text in one encoder pass, without generating anything. Weights and code (GitHub, v0.2.0) are MIT. An int8 ONNX build runs on CPU, and a vLLM plugin serves it behind POST /pooling.

The request JSON follows Jev’s shape (state, questions, criteria), but there is no /v1/systemone endpoint, so Jev clients need a small adapter.

The numbers (UnverifiedClaim). On razvilka, 735 Russian items over 15 tasks, the team reports 0.893 for FRIDA-Decisions and 0.897 for Jev on the same items (McNemar p = 0.84), ahead of migom-2b (0.853) and decider-2b (0.833). razvilka was published the same day, its scorer lives in the model’s repo, and the team ran every row. Latency: 28–34 ms per request in process on an RTX 5060 Ti; about 0.9 s on six CPU threads with the int8 build. The card has no calibration figures.

Why it matters. It is the first decision model in the catalog trained mostly on Russian (71% of 1.42M examples), and the first encoder-only one that ships a CPU build as a first-class path.

How we list it. Catalog, 5.8 maturity / 6.8 capability / 4 adoption.