GigaChat 3.1 Lightning is the compact instruct model of the GigaChat 3.1 family. It is a Mixture-of-Experts (MoE) model with 10B total parameters and 1.8B active parameters, designed for fast multilingual assistant workloads, reasoning, code, function calling, and product-style deployment.
For high-performance inference, an fp8 version of the model is available - GigaChat3.1-10B-A1.8B.
GigaChat 3.1 Lightning uses a custom MoE architecture with the following key components.
Mixture-of-Experts (MoE)
The model has 10B total parameters with 1.8B active parameters at inference time. This allows it to scale model capacity aggressively while keeping the active compute budget much lower than that of an equally large dense model.
Multi-head Latent Attention (MLA)
Instead of standard multi-head attention, the model uses MLA, which compresses the KV cache into a latent representation. This reduces memory usage and improves inference throughput, especially in long-context settings.
Multi-Token Prediction (MTP)
The model is trained with MTP, which allows it to predict multiple tokens per forward pass. In production systems, this can be used with speculative or parallel decoding techniques to improve throughput.
Training data
The base GigaChat 3 training corpus spans 10 languages and includes books, academic material, code datasets, and mathematics datasets. All data goes through deduplication, language filtering, and automatic quality checks based on heuristics and classifiers.
Synthetic data remains a major contributor to quality. Across the broader training corpus, we used approximately 5.5 trillion synthetic tokens, including:
question-answer data generated from source texts,
reverse-prompt chains for structured data generation,
model-authored notes embedded inside texts,
millions of synthetic tasks with solutions in mathematics and olympiad-style programming,
synthetic tests for code and reasoning tasks.
For the 3.1 release, we made major data improvements:
Hard-domain expansion at Stage 1.5: stronger coverage of mathematics, finance, physics, engineering, biology, chemistry, and medicine.
Stricter quality validation: our internal Revisor pipeline was extended with stronger checks for Markdown, LaTeX, and answer-format correctness.
LLM-judge validation: SFT and DPO data is validated with judges selected for the task type and response structure.
On-policy DPO data: preference pairs were generated from preview-model behavior, making them better aligned with real model failure modes.
Better product-oriented data: we expanded data for search-and-citation scenarios, file-aware code interpretation, personalization, and agentic dialogues with executable tool calls.
Improved answer style: we also revised formatting and writing guidelines to improve readability, correctness, and overall response quality.
Post-training improvements
DPO in native FP8
Unlike the preview release, GigaChat 3.1 Lightning includes a full DPO stage. This stage was redesigned for the MoE setup and trained in native FP8, not just quantized after training.
Important changes include:
MTP heads trained during DPO for better consistency between main-model predictions and MTP predictions,
weighted gamma with exponential decay over long sequences,
stronger tuning of batch size and DPO contribution,
better robustness against loop-inducing failure modes.
In our experiments, native FP8 DPO not only recovered the quality that could be lost with post-training FP8 quantization, but in some cases even exceeded the BF16 result while using substantially less memory.
Faster post-training
We also optimized the SFT pipeline with a combination of sequence packing, dynamic sequence parallelism, and additional pipeline optimizations. This reduced training cost significantly and improved GPU utilization, especially on long-context workloads.
Inference
One of the key advantages of GigaChat3.1-10B-A1.8B is its inference speed. The model (especially in MTP mode) demonstrates throughput comparable to that of significantly smaller dense models.
We measured this using vllm 0.17.1rc1.dev158+g600a039f5, concurrency=32, 1xH100 80gb SXM5.
Link to code.
Model
Output tps
Total tps
TPOT
Diff vs Lightning BF16
GigaChat-3.1-Lightning BF16
2 866
5 832
9.52
+0.0%
GigaChat-3.1-Lightning BF16 + MTP
3 346
6 810
8.25
+16.7%
GigaChat-3.1-Lightning FP8
3 382
6 883
7.63
+18.0%
GigaChat-3.1-Lightning FP8 + MTP
3 958
8 054
6.92
+38.1%
YandexGPT-5-Lite-8B
3 081
6 281
7.62
+7.5%
Benchmark Results
Domain
Metric
GigaChat-3-Lightning
GigaChat-3.1-Lightning
Qwen3-1.7B-Instruct
Qwen3-4B-Instruct
SmolLM3
gemma-3-4b-it
General
MMLU RU
0.683
0.6803
-
0.597
0.500
0.519
General
RUBQ
0.652
0.6646
-
0.317
0.636
0.382
General
MMLU PRO
0.606
0.6176
0.410
0.685
0.501
0.410
General
MMLU EN
0.740
0.7298
0.600
0.708
0.599
0.594
General
BBH
0.453
0.5758
0.3317
0.717
0.416
0.131
General
SuperGPQA
0.273
0.2939
0.209
0.375
0.246
0.201
Code
Human Eval Plus
0.695
0.7317
0.628
0.878
0.701
0.713
Total
Average
0.586
0.631
0.458
0.612
0.514
0.421
Arena Results
Arena
GigaChat-2-Lite-30.1
GigaChat-3-Lightning
GigaChat-3.1-Lightning
YandexGPT-5-Lite-8B
SmolLM3
gemma-3-4b-it
Qwen3-4B
Qwen3-4B-Instruct-2507
Arena Hard Logs V3
23.700
14.3
46.700
17.9
18.1
38.7
27.7
61.5
Validator SBS Pollux
32.500
24.3
55.700
10.3
13.7
34.000
19.8
56.100
Total Average
28.100
19.3
51.200
14.1
15.9
36.35
23.75
58.800
Usage Example
llama-cpp version>=8495 (58c81f7e8)
cmake version>=3.28.3
Build llama-cpp
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON # build for CUDA
cmake --build build --config Release --target llama-server -j 8
Start the server
shell
1# Specify the full path to your model.gguf2exportMODEL_PATH="/path/to/your/model.gguf"34./build/bin/llama-server \5 -m $MODEL_PATH\6 -np 1\7 -cb \8 -ctk q8_0 \9 -ctv q8_0 \10 -fa on \11 --n-gpu-layers 999\# number of layers on GPU12 --ctx-size 32768\13 --port 8080\14 --host 0.0.0.0 \15 --jinja
