GigaChat 3.1 Lightning
Sber's 10B mixture of experts with 1.8B active parameters, strong in Russian, with FP8, BF16 and GGUF files. MIT license.
In short. GigaChat 3.1 Lightning is a small free AI model from Sber that answers quickly, because it uses only 1.8 billion of its 10 billion parameters at a time. It is good at Russian and English, can call tools, and the MIT license allows commercial use. The 8-bit files are about 12 GB, and a 4-bit file is 6.5 GB.
Summary
GigaChat 3.1 Lightning is the compact instruct model of Sber's GigaChat 3.1 family. Sber's Habr blog announced it on 24 March 2026, together with the 700B Ultra. It is a 10B mixture of experts with 1.8B active parameters (11.48B counted on Hugging Face), multi-head latent attention and multi-token prediction. The configuration shows 26 layers, 64 routed experts with 4 used per token plus one shared, and 262,144 positions through YaRN scaling, but the card states no context length. The training corpus spans 10 languages and includes books, academic material, code and mathematics, and the card cites about 5.5 trillion synthetic tokens across the broader GigaChat 3 corpus. Version 3.1 adds a DPO stage trained in native FP8, and Sber says it fixed the repetition loops that delayed the release from January.
In the card's tables Lightning averages 0.631 across seven general and code tests, against 0.612 for Qwen3-4B-Instruct, 0.514 for SmolLM3 and 0.421 for gemma-3-4b-it. It leads on the Russian tests, with MMLU RU at 0.680 against 0.597 and RuBQ at 0.665 against 0.317 for Qwen3-4B-Instruct. Qwen3-4B-Instruct is ahead on MMLU-Pro, BBH and HumanEval Plus, and on the card's arenas Qwen3-4B-Instruct-2507 averages 58.8 against 51.2. For speed, on one H100 80 GB with vLLM at a concurrency of 32, FP8 with multi-token prediction reaches 3,958 output tokens per second against 2,866 for BF16, and the card lists YandexGPT-5-Lite-8B at 3,081. All of these numbers are Sber's own.
The FP8 weights are 12.3 GB, the BF16 weights 23.0 GB and the GGUF Q4_K_M file 6.5 GB, and the card shows Transformers, vLLM and SGLang examples with speculative decoding. The license is MIT, set in the card metadata and in Sber's announcement, and the repository has no LICENSE file. In the 30 days to 11 November 2026, Hugging Face counted about 19,000 downloads for the GGUF repository, 6,300 for BF16 and 1,900 for FP8.
Good for
- Russian-language chat, summarization and question answering on a single GPU or a well-equipped laptop.
- Fast serving, since only 1.8B parameters are active and multi-token prediction adds speculative decoding.
- Function calling in products that need a small, permissively licensed model.
- A strong small baseline for Russian-language evaluation.
Watch out for
- All 10B parameters must sit in memory even though 1.8B are active, so plan for the file sizes above.
- Sber's own tables show Qwen3-4B-Instruct ahead on several English reasoning and coding tests, and Qwen3-4B-Instruct-2507 ahead on the arenas.
- The card states no context length, and the 262,144 positions come only from the configuration.
- The repositories have no LICENSE file, so keep Sber's announcement and the card metadata with your records.
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the 10B-A1.8B GGUF Q4_K_M variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| 10B-A1.8B FP8 FP8 | 11.5B 1.8B active | 26 GB | 13 GB | 7.92 GB | Laptop or phone |
| 10B-A1.8B BF16 BF16 | 11.5B 1.8B active | 26 GB | 13 GB | 7.92 GB | Laptop or phone |
| 10B-A1.8B GGUF Q4_K_M INT4 | 11.5B 1.8B active | 26 GB | 13 GB | 7.15 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- Best small open models you can run on a laptop
- Sber's open models: a 432B chat model, speech and embeddings
Frequently asked questions
- What is GigaChat 3.1 Lightning?
- GigaChat 3.1 Lightning is a small free AI model from Sber that answers quickly, because it uses only 1.8 billion of its 10 billion parameters at a time. It is good at Russian and English, can call tools, and the MIT license allows commercial use. The 8-bit files are about 12 GB, and a 4-bit file is 6.5 GB.
- Can I use GigaChat 3.1 Lightning commercially?
- Yes. GigaChat 3.1 Lightning is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is GigaChat 3.1 Lightning open source?
- GigaChat 3.1 Lightning is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run GigaChat 3.1 Lightning on my own computer?
- Yes. The smallest variant, 10B-A1.8B GGUF Q4_K_M, needs about 7.15 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was GigaChat 3.1 Lightning released?
- GigaChat 3.1 Lightning was released by Sber (GigaChat) on March 24, 2026.
- How long a context window does GigaChat 3.1 Lightning have?
- GigaChat 3.1 Lightning supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- GigaChat 3.1 Lightning model card (Hugging Face) huggingface.co
- GigaChat-3.1: a big update to the big models (Sber, Habr, 24 March 2026) habr.com
- GigaChat 3.1 Lightning GGUF files (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| GigaChat 3.5 Ultra Sber (GigaChat) | 434B – 438B 28B active | 256K | MIT (Permissive) | 299 GB Big-memory desktop | |
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Nemotron 3 Nano 4B NVIDIA | 3.97B | 256K | NVIDIA Nemotron (Permissive) | 2.74 GB Laptop or phone |