YandexGPT 3
Yandex · Russia
- Params
- 30B
- Context
- 8,192
- Min VRAM
- ,
- GGUF
- No
Strong Russian-language generation, integration with Yandex services, fast inference via Yandex Cloud.
Region slug: russiaRussia / CIS
AI models, vendors, and hardware buying realities from Russia / CIS. Curated by myaihardware.com.
21
Models
8
Vendors
5
Languages
Why Russia / CIS matters for AI builders
Russia matters because it holds roughly 10% of the world's semiconductor-grade neon and rare-earth metals, creating a supply-chain lever that AI hardware builders cannot ignore. Its dominant model families, like Yandex's YaLM and Sber's Kandinsky-2, are optimized for Russian-language and low-latency inference on older GPU stacks, offering a real-world stress test for hardware efficiency. The underrated angle: Russia's vast, underutilized hydroelectric and nuclear grid capacity in Siberia makes it a potential site for low-cost, low-carbon AI datacenter buildouts, sidestepping Europe's energy constraints.
Market context
Post-2022 export controls cut Russian commercial labs off from H100/H200 supply, pushing Yandex and Sber to lean on stockpiled A100s, Huawei Ascend, and continual-pretrain shortcuts that piggyback on Qwen and Llama base weights. The commercial market is a Yandex-Sber duopoly: YandexGPT 5 Pro and GigaChat 2 Max both target GPT-4o parity on Russian content via cloud-only APIs that western customers cannot legally consume due to SDN-list status. Beneath that, a vibrant community scene around Vikhr, Saiga and T-Bank's Apache-2.0 releases has made Russian one of the best-served non-English languages on Hugging Face, while Kazakh (KazLLM), Ukrainian (UA-LLM, ukrLM) and Belarusian (BelarusBERT) projects address each post-Soviet language with state and academic backing where Yandex/Sber's Russocentric output would not fit.
Models (21 of 21)
Filter by vendor, license, or parameter range.
Restricted
YandexGPT 3 Pro
Yandex · Russia
- Params
- 70B
- Context
- 32,768
- Min VRAM
- ,
- GGUF
- No
High quality Russian text generation, good factual accuracy for Russian content, long context.
Region slug: russiaYandexGPT 3 Lite
Yandex · Russia
- Params
- 7B
- Context
- 4,096
- Min VRAM
- ,
- GGUF
- No
Low latency, cost-efficient for simple tasks, well-tuned for Russian language.
Region slug: russiaYandexGPT 5 Pro
Yandex · Russia
- Params
- 200B
- Context
- 65,536
- Min VRAM
- ,
- GGUF
- No
Best-in-class Russian reasoning, function calling, agentic flows via Yandex Cloud Foundation Models.
Region slug: russiaGigaChat Pro
Sber · Russia
- Params
- 29B
- Context
- 32,768
- Min VRAM
- ,
- GGUF
- No
Flagship Sber model with strong Russian comprehension, tool use, retrieval; tightly integrated with Sber ecosystem (SberCloud, SaluteSpeech).
Region slug: russiaGigaChat Lite
Sber · Russia
- Params
- 7B
- Context
- 8,192
- Min VRAM
- ,
- GGUF
- No
Cheap, fast tier of GigaChat; suitable for classification, summarization, light chat.
Region slug: russiaGigaChat 2 Max
Sber · Russia
- Params
- 70B
- Context
- 131,072
- Min VRAM
- ,
- GGUF
- No
Russian SOTA reasoning model, multimodal (text + image) and agentic tool-calling; Sber's GPT-4o competitor.
Region slug: russiaruGPT-3.5 13B
Sber AI Foundation Models (SberDevices) · Russia
- Params
- 13B
- Context
- 2,048
- Min VRAM
- 16 GB
- GGUF
- Yes
Open weights, MIT-licensed; the base model behind GigaChat first-gen and many community Russian finetunes.
Region slug: russiaT-Pro-IT 1.0 (32B)
T-Bank (formerly Tinkoff) · Russia
- Params
- 32B
- Context
- 32,768
- Min VRAM
- 24 GB
- GGUF
- Yes
Open-weight Russian SOTA at 32B; tops MERA among open models; commercial-friendly Apache license.
Region slug: russiaT-Lite-IT 1.0 (8B)
T-Bank (formerly Tinkoff) · Russia
- Params
- 8B
- Context
- 8,192
- Min VRAM
- 8 GB
- GGUF
- Yes
Best 8B Russian model end of 2024, Apache 2.0, easy single-GPU inference.
Region slug: russiaMTS Cotype (Nano + Pro)
MTS AI · Russia
- Params
- 8B
- Context
- 32,768
- Min VRAM
- 8 GB
- GGUF
- Yes
MTS-built Russian assistant family; Nano open-weight 1.5B/8B variants compete with T-Lite; integrates with MTS AI Marketplace.
Region slug: russiaVikhr-Nemo-12B-Instruct
Vikhr Team (community) · Russia
- Params
- 12B
- Context
- 131,072
- Min VRAM
- 12 GB
- GGUF
- Yes
Community SOTA for 12B Russian; 128K context from Mistral Nemo base; agent / RAG ready.
Region slug: russiaVikhr-7B-Instruct
Vikhr Team (community) · Russia
- Params
- 7B
- Context
- 8,192
- Min VRAM
- 6 GB
- GGUF
- Yes
First widely-used community Russian Mistral finetune; runs on consumer 12GB GPU; permissive license.
Region slug: russiaVikhr-Llama-3.2-1B-Instruct
Vikhr Team (community) · Russia
- Params
- 1B
- Context
- 131,072
- Min VRAM
- 2 GB
- GGUF
- Yes
Edge-class Russian model, CPU / Jetson deployable; useful for embedded Russian chat.
Region slug: russiaSaiga-Llama3-8B
Ilya Gusev (community) · Russia
- Params
- 8B
- Context
- 8,192
- Min VRAM
- 8 GB
- GGUF
- Yes
Most-downloaded Russian community chat model on HF; well-documented training recipe; baseline used by RuArena.
Region slug: russiaSaiga-Mistral-7B (Saiga 2)
Ilya Gusev (community) · Russia
- Params
- 7B
- Context
- 8,192
- Min VRAM
- 6 GB
- GGUF
- Yes
Iconic 2023-era Russian chat model; reproducible LoRA recipe popularized RU finetuning.
Region slug: russiaKazLLM-8B
IssAI (Institute of Smart Systems & AI, Nazarbayev University) · Kazakhstan
- Params
- 8B
- Context
- 8,192
- Min VRAM
- 8 GB
- GGUF
- Yes
First state-backed Kazakh foundation model; trained on 148B tokens (40B Kazakh, 30B Russian, 78B English); covers Cyrillic + Latin Kazakh scripts.
Region slug: russiaKazLLM-70B
IssAI (Institute of Smart Systems & AI, Nazarbayev University) · Kazakhstan
- Params
- 70B
- Context
- 8,192
- Min VRAM
- 48 GB
- GGUF
- Yes
Largest open Kazakh-first model; multilingual KK/RU/EN; backed by Kazakhstan Ministry of Digital Development.
Region slug: russiaBelarusBERT
KassymKulov / Belarusian AI Community · Belarus
- Params
- 110M
- Context
- 512
- Min VRAM
- 1 GB
- GGUF
- No
Only widely-available Belarusian encoder; useful for NER, classification, POS-tagging on bel-Cyrl text.
Region slug: russiaUA-LLM (UA-Llama-3-8B)
ITU Kyiv / Ukrainian Catholic University community · Ukraine
- Params
- 8B
- Context
- 8,192
- Min VRAM
- 8 GB
- GGUF
- Yes
Best open Ukrainian-language chat model; trained on lang-uk's UA corpus; community-driven post 2022.
Region slug: russiaukrLM-base
lang-uk / Ukrainian Catholic University · Ukraine
- Params
- 360M
- Context
- 512
- Min VRAM
- 1 GB
- GGUF
- No
Reference Ukrainian encoder for NER, classification, embeddings; backbone of UA-MERA evaluation.
Region slug: russiaNotable vendors
YandexSberT-BankMTS AIVikhr TeamIlya GusevIssAIlang-uk
Run benchmarks
llama.cpp, MLPerf, and inference leaderboards.
VRAM calculator
Will this model fit on your GPU?
Hardware guides
Local LLM setup, homelab, mini-PC builds.
Methodology
How we benchmark, verify ASINs, and source data.
Beginner's guide
New to local AI? Start here.
Sourcing & commerce disclosure
Model metadata on this page is compiled from each vendor's public HuggingFace repo, model card, or research paper at time of indexing. Performance figures (KMMLU, MMLU, HAERAE, AraBench, IndicGEM, etc.) link directly to source benchmarks where available; missing or partial values are shown as ", ".
Outbound hardware links (GPUs, mini-PCs, accelerators) are Amazon Associate links using store ID fredoline-20. As an Amazon Associate we earn from qualifying purchases. Prices and availability shown elsewhere on the site are snapshots, not live quotes — verify on Amazon before purchasing. See About / disclosures for the full policy.