GigaChat3.1-10B-A1.8B

GigaChat3.1-10B-A1.8B, or GigaChat 3.1 Lightning, is a compact instruction-tuned model from the GigaChat 3.1 family, built on a Mixture-of-Experts (MoE) architecture. With a total of 10 billion parameters, only 1.8 billion are activated per token during computation, dramatically reducing memory footprint and speeding up inference without sacrificing quality. The model’s architecture is based on Multi-head Latent Attention (MLA), which compresses the KV-cache into a latent representation, thereby saving memory when processing long contexts. Additional performance gains come from Multi-Token Prediction (MTP), a mechanism that allows the model to predict several tokens in a single forward pass.

A key difference from GigaChat3-10B-A1.8B is the alignment stage (DPO), conducted entirely in native FP8 using the DeepGEMM library, rather than applying post-hoc compression such as quantization. This not only preserved but in several cases exceeded the quality of BF16 training while halving memory consumption.

The model was trained on a corpus covering 10 languages, including books, scientific materials, code, math datasets, and approximately 5.5 trillion synthetic tokens. For version 3.1, a large-scale data improvement effort was undertaken: coverage of complex domains (mathematics, physics, biology, chemistry, finance) was expanded, validation was tightened through an internal Revisor pipeline and an LLM-as-a-judge system, and preferences were added based directly on “live” behavior from previous model versions (on-policy DPO). This approach ensures high relevance, accuracy, and resistance to repetition.

Thanks to its high speed, compactness, and quality, GigaChat 3.1 Lightning is ideally suited for building fast multi-task assistants that can work with long documents, explain complex concepts, and generate and debug code. Key use cases include: customer support automation with function calling, structured content generation (markdown, LaTeX), personalized recommendations, search-and-citation dialogues, executable tool calls in assistants, and any task where low latency and high answer accuracy are critical — chatbots, RAG systems, and on-device AI assistants for resource-constrained environments.

Announce Date: 21.03.2026
Parameters: 12B
Experts: 64
Activated at inference: 2B
Context: 263K
Layers: 26
Attention Type: Multi-head Latent Attention
Developer: Sber AI
Transformers Version: 4.53.2
License: MIT

Use our pre-built public endpoints for free to test inference and explore GigaChat3.1-10B-A1.8B 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.