GigaChat opens Russian MoE weights

The important story here is not that GigaChat suddenly beats frontier Western or Chinese models — it doesn’t — but that Russian-language LLM work finally has a more serious open reference point with weights, benchmarks, and documentation.

• –The paper explicitly frames GigaChat as a family of Russian LLMs with both base and instruction-tuned variants, plus three open models released for research and industrial use
• –The Hugging Face model card shows a DeepSeek-style MoE implementation path and custom code, which makes the community architecture debate understandable even if the release is positioned as its own model family
• –Reported results look strongest on Russian-focused evaluation, especially MERA, while coding and English benchmarks still trail the closed GigaChat tiers and stronger global competitors
• –Outside commentary has already called out the gap between the open-weight model and top open alternatives like Qwen and Llama, so the value here is ecosystem coverage, not frontier leadership
• –For developers working on Russian NLP, this is still useful: MIT-licensed weights, vLLM examples, long context, and public benchmarks are much more actionable than vague API-only claims