GigaChat3-702B-A36B-preview

GigaChat 3 Ultra Preview is a massive MoE model with 702 billion total parameters, of which only 36 billion are active in generating each token. A key feature is the use of Multi-head Latent Attention (MLA) — a mechanism that compresses the Key-Value (KV) cache into a latent vector. For users, this means the ability to work with massive contexts and long documents without an exponential increase in GPU memory requirements. The model is trained using a Multi-Token Prediction (MTP) objective — unlike classical solutions that predict only the next word, this model can predict multiple tokens in a single forward pass. This implements a native speculative decoding mechanism, accelerating response generation by up to 40% without quality loss. Technically, this makes it one of the fastest models in the "heavy" LLM class.

Unlike many Russian models that are the result of fine-tuning foreign counterparts, GigaChat 3 was trained from scratch on a unique dataset of over 20 trillion tokens. This dataset includes a vast amount of Russian-language sources and adds languages rarely featured in model training — from Chinese and Arabic to Uzbek and Kazakh. This provides the model with a deep understanding of the Russian language and Russian cultural context, which are inaccessible to foreign LLMs. The model leads in both Russian and international benchmarks: MERA (Industrial) - 0.824, HumanEval+ (Code) - 0.8659, GSM8K (Math) - 0.9598, etc. In terms of quality, GigaChat 3 Ultra confidently outperforms GigaChat 2 Max on all key benchmarks.

The model is optimized for On-premise and Private Cloud installations in large enterprises where data security is critical (air-gapped environments without internet access). It is ideally suited for complex analytics of large document arrays (RAG), automation of L2/L3 technical support (requiring deep contextual understanding), programmer assistants, and code generation within closed corporate repositories. The model supports popular inference engines (vLLM, SGLang, LMDeploy, TensorRT-LLM) and operates in BF16 and FP8 modes for optimal performance.

Announce Date: 19.11.2025
Parameters: 715B
Experts: 256
Activated at inference: 36B
Context: 131K
Layers: 64
Attention Type: Multi-head Latent Attention
VRAM requirements: 344.5 GB using 4 bits quantization
Developer: Sber AI
Transformers Version: 4.53.2
License: MIT

Use our pre-built public endpoints for free to test inference and explore GigaChat3-702B-A36B-preview capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.