The Pulse
Sber Releases 432B-Parameter GigaChat 3.5 Reasoning Model
Sber has released GigaChat 3.5 Reasoning, a 432-billion-parameter mixture-of-experts model with 28 billion active parameters. The open-weight model uses online reinforcement learning, supports a 262,000-token context window, and is distribu
AI.info Team ·
Sber released GigaChat 3.5 Reasoning on September 10, presenting it as the first GigaChat model trained with full online reinforcement learning and making its weights and launch code available to developers. The company’s release describes a 432-billion-parameter mixture-of-experts model with 28 billion active parameters.
“Мы выпускаем GigaChat 3.5 Reasoning, первую модель GigaChat с полноценным рассуждением, обученную на технологии online RL,” Emil Shakirov, head of Alignment at GigaChat, wrote in Sber’s company blog on Habr. The English translation is: “We are releasing GigaChat 3.5 Reasoning, the first GigaChat model with full reasoning, trained with online RL.”
«Мы выпускаем GigaChat 3.5 Reasoning, первую модель GigaChat с полноценным рассуждением, обученную на технологии online RL.»
Emil Shakirov, head of Alignment, GigaChat at Sber
Six specialists become one model
Sber says the post-training process began with a supervised fine-tuning checkpoint and split into six separate experts. The specialists covered STEM tasks, conventional code generation and editing, repository-level code agents, general agents, dialogue, and instruction following with structured output and long-context tasks.
Each expert used its own reward design. Mathematical answers could be checked against a known result, code could be executed, repository patches could be tested, and agents could be judged by the final state of an environment. For dialogue, Sber used side-by-side evaluation by a language-model judge, while instruction-following tasks relied on format and constraint checks.
The team trained the experts with CISPO, an optimization method related to GRPO. Sber says CISPO keeps learning signals for rare tokens that can mark a change in direction during reasoning, such as checking an intermediate result or backing out of a mistake. After separate reinforcement-learning runs, the six experts were combined into one release model through on-policy distillation.
A hybrid architecture built for long context
GigaChat 3.5 Reasoning combines Multi-head Latent Attention with GatedDeltaNet linear-attention layers. The model also includes GatedNorm and three multi-token-prediction heads intended to accelerate speculative decoding.
The published model card lists a maximum context length of 262,000 tokens. It provides an FP8 inference repository, while a separate BF16 repository is intended for fine-tuning and custom quantization. Both repositories are published under the MIT license on Hugging Face and a separate BF16 model page.
Sber’s recommended serving path uses SGLang with tensor and expert parallelism. The company’s instructions describe an eight-H100 configuration for the FP8 model and a two-node, 16-GPU setup for the BF16 weights. The serving configuration exposes the reasoning trace in a separate reasoning_content field and supports function-call parsing.
Large gains over GigaChat 3.5 Instant
Sber reports marked improvements over GigaChat 3.5 Instant on several internal evaluations. GPQA-Diamond rises from 61.11 to 82.32, AIME 2026 at mean@32 increases from 67 to 92, and IFBench with a loose prompt moves from 43.66 to 77.
The model card places GigaChat 3.5 Reasoning close to DeepSeek V4 Flash Preview on the company’s combined evaluation table, although the results are not uniformly better. GigaChat scores higher on IFBench and structured output, while DeepSeek leads on several agentic coding and mathematics tests. On SWE-bench Verified, GigaChat records 64.7 compared with DeepSeek’s 78.6; on Terminal-Bench 2, the figures are 30.3 and 56.6.
Sber also reports a lower reasoning-token budget on difficult mathematics tasks. Across AIME 2025, AIME 2026, HMMT, and IMOAnswerBench, the company says GigaChat 3.5 Reasoning uses 37% fewer reasoning tokens than DeepSeek V4 Flash Preview in the reported samples. The underlying figures range from a 23% reduction on AIME 2026 to 41% on IMOAnswerBench.
Training required more than nine months
The release describes online reinforcement learning at this scale as an infrastructure problem as much as a model-training problem. Sber says the project took more than nine months, with engineers working to keep training and inference engines aligned across a mixture-of-experts router that can select different experts for the same input.
The team used rollout routing replay to preserve the expert choices made during inference and reuse them during training. Sber says the method reduced the difference between the two engines by roughly an order of magnitude in its internal measurements. An asynchronous training pipeline, meanwhile, separated teacher and rollout GPU pools and ran more than 2.5 times faster than the starting implementation.
Weight synchronization fell from several minutes to eight seconds per training step after Sber adapted an implementation from the Miles repository for the Verl training framework. The company also says native multi-token prediction added about 30% to training speed, while FP8 training produced a 1.3-to-2-times speed increase compared with its earlier setup without a measured quality drop against BF16.
Open weights, with local deployment requirements
Users can try the model through the GigaChat service, while developers can download the weights and run inference locally. The hardware requirements are substantial: Sber’s own launch examples call for eight H100 GPUs for FP8 inference, or a larger 16-GPU, two-node configuration for the BF16 checkpoint.
Sber has not presented the release as a small local model. Its significance is different: the company is publishing a full reasoning model with a large total parameter count, separate domain experts during training, tool-use support, structured output, and a long context window. The open repositories give researchers access to the weights and serving path, but practical deployment remains a datacenter-scale task.
The release is dated September 10, 2026. Sber’s announcement says an API offering for businesses is planned, while the public model repositories and the GigaChat service provide the immediate access points.