Training a reasoning model via online RL took Sber's team more than nine months. The effort culminated in the release of GigaChat 3.5 Reasoning, the first Russian model featuring a fully fledged chain-of-thought mechanism. Sber released the weights and inference code openly on Hugging Face, providing a local alternative to foreign reasoning systems like OpenAI o1 or DeepSeek-R1.
Six-expert architecture
Standard Supervised Fine-Tuning (SFT) trains a network to mimic the format of a correct answer, but leaves fatal blind spots in rigorous proof logic. To move beyond mere pattern memorization, the team passed the post-SFT model entirely through an online RL loop. Engineers branched six independent domain experts out of the base checkpoint—each trained against its own reward function—before merging them back into a single monolithic model.
"After SFT, the model went entirely through online RL. Not just a single domain and not a single final stage, but six distinct experts (math, code, agents, and others), each with its own reward function, which were later assembled back into one model."
Reassembling the experts was achieved through on-policy distillation. Each domain was trained using the CISPO algorithm from the MiniMax-M1 paper—a close relative of GRPO. The key difference: CISPO preserves tokens of doubt and self-correction ("let's double-check", "wait, there's an error here") instead of clipping them when out of distribution, smoothly scaling down the gradient penalty instead.
Curriculum dynamics and benchmarks
Reasoning model training heavily relies on difficulty filtering: tasks that the current checkpoint solves with a probability above 75% are immediately dropped from the pipeline due to their near-zero training gradient.
Adopting step-by-step reasoning sharply reduced hallucinations in code, rigorous math, and agentic workflows. Compared to the base GigaChat 3.5 Instant, the GPQA-Diamond benchmark climbed from 61.11 to 82.32 points, the AIME-2026 score (mean@32) rose from 67 to 92, and the IFBench Loose Prompt test jumped from 43.66 to 77.00.
For the enterprise sector, open-sourcing GigaChat 3.5 Reasoning fundamentally shifts deployment economics: businesses get an autonomous B2B agent reasoning model that can run strictly on-premise, without sending sensitive corporate data to foreign cloud APIs.