New Delhi, Sept. 17 -- The open-weight Mixture-of-Experts model combines six specialised reasoning experts, hybrid attention and a 262,000-token context window for coding, agentic and analytical workloads.

Sber has introduced GigaChat 3.5 Reasoning, an artificial intelligence model designed to plan, verify and work through complex, multi-stage tasks. The release adds a reasoning mode to the GigaChat family and gives developers access to model weights for testing, fine-tuning and product integration.

Announced on September 10, 2026, the model can divide problems into stages, identify missing information, use external tools and check intermediate results before responding, according to the company. It can also revisit earlier decisions when a proposed solution does not satisfy the task's requirements.

Sber said the model recorded its largest improvements over GigaChat 3.5 Ultra Instruct in mathematics, coding, instruction following and structured output.

Reasoning for coding and agentic systems

The model is aimed at tasks where the quality of an answer depends on several connected decisions. These include debugging software, examining documents for conflicting provisions, preparing schedules within budget and resource constraints, and handling customer requests that require information from multiple systems.

Software development is one of its primary applications. The model has been trained on algorithms, code editing, test generation and repository-level assignments in which proposed patches are applied and tested. This allows developers to evaluate whether it can reason across code dependencies instead of treating each fragment in isolation.

In document analysis, the model can compare requirements across contracts and identify potentially inconsistent clauses for further review. Planning applications could include building routes or schedules while accounting for deadlines, budgets and available resources.

Customer-support systems could use the model to clarify requests and consult different sources before producing a response. Analytical workflows could similarly combine calculations with successive checks of a hypothesis. Their effectiveness in production will depend on tool access, data quality, evaluation frameworks and safeguards.

Anton Frolov, senior vice president and head of GenAI Development at Sberbank, said the ability to identify missing information, validate intermediate conclusions and modify a proposed solution is particularly important in programming, analytics and agent-based applications.

Online reinforcement learning with six experts

GigaChat 3.5 Reasoning is the first model in the GigaChat family to feature full reasoning trained through online reinforcement learning, according to its technical documentation.

Its post-training process uses six independently trained domain experts covering science and mathematics, coding, code agents, general agents, dialogue, and capabilities including instruction following, long-context processing and structured output.

Each expert uses feedback suited to its domain. Mathematical outputs are checked against final answers, coding tasks through execution and repository patches through automated tests. General agents are assessed against the final environment state, while dialogue responses are compared by an AI judge.

The experts are subsequently combined through on-policy distillation. Under this approach, the main model generates its own solution path while the relevant expert provides token-level supervision.

Hybrid architecture and developer access

The model uses a Mixture-of-Experts architecture containing 432 billion parameters, of which 28 billion are active during processing. It supports a maximum context length of 262,000 tokens, allowing it to work with substantial documents, extended conversations and large collections of code.

Its hybrid architecture combines Multi-head Latent Attention with GatedDeltaNet linear-attention layers. The model was trained natively in FP8 precision, while a BF16 version is available for fine-tuning and custom quantisation. Three multi-token prediction heads are included to support speculative decoding.

Sber reported that the model's IFBench score increased from 44 to 77 compared with the non-reasoning version. Its Natural Plan score rose from 64 to 80, while its Live Code Bench v6 score increased from 56 to 85.

The company also claimed that the model used an average of 37% fewer tokens on mathematical problems than DeepSeek V4 Flash Preview. The comparison relates to a specific workload and does not necessarily translate into an equivalent reduction in operating costs or response times.

Users can activate reasoning mode for more complex requests, while straightforward questions can receive immediate responses. Sber said difficult problems may require tens of thousands of tokens, leaving enterprises to balance accuracy against processing time and resource consumption.

GigaChat 3.5 Reasoning is available to users and developers, with its model weights released on Hugging Face under the MIT licence. Sber has also said that API access for businesses will follow.

The MIT-licensed release allows engineering teams to benchmark the model on their own codebases and agent workflows. Production viability will depend on hardware requirements, inference efficiency and how consistently the model performs outside Sber's reported evaluations.

Published by HT Digital Content Services with permission from TechCircle.