Sber Opens GigaChat 3.5 Reasoning as Russia’s First Open Reasoning Model
Sber published GigaChat 3.5 Reasoning on September 10, 2026, describing it as the first GigaChat model with a full reasoning mode trained via online reinforcement learning, with weights available on Hugging Face under an open license. Architecture notes describe a roughly 432-billion-parameter mixture-of-experts design with about 28 billion active parameters per query, plus tooling for SGLang and related serving stacks that separate reasoning traces from final answers.
Filed under Product and dated September 11, 2026, this AI4Russia briefing treats vendor benchmarks as directional. Sber materials claim gains versus prior Ultra instruct variants on coding, planning, and math suites, and position the release as Russia’s first open model with comparable reasoning behaviour. Consumer access via giga.chat is already highlighted; business API timing remains a follow-on promise rather than a finished enterprise SLA.
Why it matters: procurement and engineering teams in Moscow watching draft sovereign-model preferences need practical options they can evaluate offline. An open reasoning checkpoint changes the conversation from closed API demos to reproducible local tests—provided organisations have the GPUs, evaluation harnesses, and legal clarity to run them.
What it means in practice
Russian operators should inventory which workflows need multi-step reasoning versus simple drafting; confirm lawful data access; assign a human owner; and run time-boxed pilots with logging and override rights. Prefer serving setups that expose reasoning traces for audit. Treat MIT-style openness as helpful for inspection, not as a guarantee of production safety or regulatory fitness under evolving sovereign-model rules.
Caveats come first. Large MoE checkpoints are expensive to host; vendor leaderboards can overstate field reliability; and draft government preference rules may still steer buyers toward labelled “sovereign” stacks regardless of open weights. AI4Russia presents the release as product context only.
What to watch next: independent evaluations on Russian-language enterprise tasks; API pricing and latency for business users; and how GigaChat 3.5 Reasoning is classified under secondary rules due through 2027. Readers can continue on the AI4Russia homepage for related stories, or browse the Newsroom for additional briefings.
Bottom line: treat this update as orientation, not instruction. Domestic model releases in Russia are accelerating and uneven. Organizations that benefit most will measure pilots honestly, keep people accountable, and avoid locking critical workflows to a single stack before rules and hardware capacity stabilise.