On September 10, 2026, Sber released GigaChat 3.5 Reasoning — Russia's first open reasoning model trained from scratch, rather than fine-tuned from an existing open model. The 432B-A28B MoE weights in FP8 format have been published on Hugging Face under the MIT license and are permitted for commercial use.
What happened
The model's post-training followed an online RL pipeline: first, six domain experts were trained — math/STEM, code, code agent, general agent, dialogue, and instruction following, each with its own reward function — and then they were combined into a single model through on-policy distillation (OPD), with CISPO serving as the RL algorithm. Compared to GigaChat 3.5 Instant and S3.5 Ultra Instruct, the following results are claimed: SWE-bench Verified — from 42.6 to 64.7, AIME-2026 (mean@32) — from 67 to 92, IFBench — from 43.66 to 77, GPQA-Diamond — from 61.11 to 82.32. On complex math (AIME 2025 and 2026, HMMT, IMOAnswerBench), the model uses on average approximately 37% fewer reasoning tokens. Reasoning is already available to end users in GigaChat's reasoning mode.
Context
The 432B-A28B format means mixture-of-experts: out of 432 billion total parameters, only 28 billion are active per token, which reduces inference cost while maintaining model capacity. The architecture additionally combines efficiency techniques: hybrid attention MLA + GatedDeltaNet, GatedNorm, three MTP heads for speculative decoding, and a context of up to 262K tokens. Open weights under the MIT license grant the right to freely download, deploy, and monetize the model, and the publication of the post-training description (without pre-train details) is a rare case in the industry of a public breakdown of an RL pipeline.
Why this matters for the industry
For the market, this is the first Russian open reasoning artifact that can be deployed on its own inference without being tied to someone else's API: startups and companies can deploy, fine-tune, and sell reasoning products on an MIT model without licensing risks. An open alternative creates price pressure on vendors monetizing access to closed reasoning APIs — the buyer can now compare the cost of an API request with the price of a self-hosted model. The published pipeline (separate domain rewards, on-policy distillation, CISPO) becomes a public example for other teams, and with independent benchmark verification, the model could become an open baseline for reasoning tasks around which fine-tunes and derivatives will be built.
Why this matters for users
End users get the model immediately, in GigaChat's reasoning mode, without needing to deploy anything. For developers, the weights are already on Hugging Face in the ai-sage/GigaChat3.5-432B-Reasoning repository, and the official path for self-launch is through SGLang with a provided example on 8x H100. For teams with a ready 8-GPU node, this is a working solution for complex code and math tasks on their own infrastructure, where data doesn't leave the company's perimeter, and complex requests with long reasoning are cheaper.
What is still unknown / limitations
There are no public measurements of latency, throughput, or request cost yet, so the actual inference cost needs to be evaluated independently. All benchmarks are claimed relative to Sber's own models, the absolute level of GigaChat 3.5 Reasoning compared to other frontier models is not disclosed in the materials, and statements about parity with the strongest models do not name specific competitors. Support for launching in SGLang at the time of release is formatted as pull request sgl-project/sglang#29189 and has not yet entered a stable release, and self-reported figures — especially AIME-2026 and GPQA-Diamond — await independent reproduction.
Sources
- Hugging Face — official model card ai-sage/GigaChat3.5-432B-A28B-Reasoning
- Habr — Sber's article "GigaChat 3.5 Reasoning — Russia's first open reasoning model"
Author
Look at AI, editorial team