Local AI models: which to run on your own computer or server
Local AI models run on your computer or server, and your data goes nowhere. On 16 October 2026, a laptop with 16 GB of memory fits gpt-oss-20b and the small Gemma 4 models, a workstation fits Qwen3.8–27B and Gemma 4 31B, and a server fits gpt-oss-120b, DeepSeek, GLM, and Kimi. The easiest ways to run them are Ollama, LM Studio, or Jan.
In short
- Laptop with 16 GB of memory: gpt-oss-20b or Gemma 4 E4B
- Workstation: Qwen3.8–27B or Gemma 4 31B
- Server with one 80 GB graphics card: gpt-oss-120b
- Several servers: DeepSeek-V4.1-Flash, GLM-5.3, Kimi K3
- Run in a window, without the command line: LM Studio or Jan
- Run with one command: Ollama
- A model for the whole department: vLLM
The list
Order: Arena Text, checked 16 October 2026.
- #1
Kimi
Moonshot AI, China
- What it is good for
- A very large model with a long context for a server cluster, with a license that has conditions for large services.
- Why it is on the list
- Place 18 on Arena Text as of 16 October 2026, model Kimi K3 (Max)
- Price
- see the plans on the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- kimi.com
- Glossary term
- Kimi
- #2
Qwen
Alibaba (Qwen team), China
- What it is good for
- An Alibaba family from 27 billion parameters for a workstation to a flagship for a server cluster.
- Why it is on the list
- Place 24 on Arena Text as of 16 October 2026, model qwen3.8-max
- Price
- see the plans on the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- chat.qwen.ai
- Glossary term
- Qwen
- #3
GLM (Z.ai)
Z.ai (Zhipu AI), China
- What it is good for
- Open Z.ai models for code and agents, small versions under MIT.
- Why it is on the list
- Place 28 on Arena Text as of 16 October 2026, model glm-5.3-max
- Price
- API: GLM-5.3-Flash from $0.15 per 1 million input tokens, GLM-5.3 from $1.4; chat price not checked (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- chat.z.ai
- Glossary term
- GLM (Z.ai)
- #4
DeepSeek
DeepSeek, China
- What it is good for
- Large open models under MIT for a server: texts, code, reasoning.
- Why it is on the list
- Place 38 on Arena Text as of 16 October 2026, model deepseek-v4.1-flash-max
- Price
- chat and app free; API from $0.15 per 1 million input tokens (DeepSeek-V4.1-Flash, off-peak hours) (16 October 2026) plans
- Official site
- chat.deepseek.com
- Glossary term
- DeepSeek
- #5
Gemma
Google DeepMind, USA
- What it is good for
- Open Google models from phones to workstations, with the Apache 2.0 license for Gemma 4.
- Why it is on the list
- Place 76 on Arena Text as of 16 October 2026, model gemma-4-31b
- Price
- free (open weights) (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- ai.google.dev/gemma
- Glossary term
- Gemma
- #6
Vibe (formerly Le Chat)
Mistral AI, France
- What it is good for
- European models: some under Apache 2.0, with the flagship’s weights promised by the end of October 2026.
- Why it is on the list
- Place 111 on Arena Text as of 16 October 2026, model mistral-large-4
- Price
- free plan; Pro $14.99 per month (16 October 2026) plans
- Official site
- chat.mistral.ai
- Glossary term
- Vibe (formerly Le Chat)
- #7
gpt-oss
OpenAI, USA
- What it is good for
- An open OpenAI model in two sizes: for a laptop with 16 GB of memory and for a server with one graphics card.
- Why it is on the list
- Place 215 on Arena Text as of 16 October 2026, model gpt-oss-120b
- Price
- free (open weights) (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- openai.com/index/introducing-gpt-oss
- #8
Llama*
Meta, USA
- What it is good for
- Open Llama 4 models with a long context; a license with restrictions and a note on Meta’s status in Russia.
- Why it is on the list
- Place 242 on Arena Text as of 16 October 2026, model llama-3.1-405b-instruct-bf16
- Price
- free (open weights, download on request) (16 October 2026) plans
- Russian language
- no per the site
- Official site
- dev.meta.ai/llama
- Glossary term
- Llama
- #9
Ollama
Ollama, USA
- What it is good for
- Installs and runs a model with one command on Mac, Windows, and Linux.
- Why it is on the list
- According to Ollama, as of October 2026 the program is installed more than 9 million times a month, and more than 1 billion models have been downloaded through it. Ollama, October 2026
- Price
- free for local use; cloud models: free plan with pay-as-you-go usage, Pro $20 per month, Max $100 per month (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- MIT
- Official site
- ollama.com
- Glossary term
- Ollama
- #10
LM Studio (LM Studio Bionic)
Element Labs, USA
- What it is good for
- A chat window and a model catalog, a Russian interface, and a local server for your own programs.
- Why it is on the list
- On 16 July 2026, LM Studio released Bionic, an agent for open models that runs locally or in the cloud on servers in the USA. LM Studio, 16 July 2026
- Price
- free for local models; cloud models from $20 per month (Bionic+), Pro $100 per month (16 October 2026) plans
- Russian language
- yes per the site
- Official site
- lmstudio.ai
- Glossary term
- LM Studio (LM Studio Bionic)
- #11
llama.cpp
ggml-org (llama.cpp team and Hugging Face), open-source project
- What it is good for
- The engine that many launch programs run on; it works even on a CPU without a graphics card.
- Why it is on the list
- The llama.cpp project was opened on 10 March 2023; as of 16 October 2026 it has 131 thousand stars on GitHub. GitHub, 16 October 2026
- Price
- free (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- MIT
- Official site
- github.com/ggml-org/llama.cpp
- #12
vLLM
open-source project (started at Sky Computing Lab, UC Berkeley), USA
- What it is good for
- A model server for a department or a company: many users on one graphics card.
- Why it is on the list
- In the vLLM paper of 12 September 2023, the authors showed a 2–4 times higher throughput for popular models at the same latency. arXiv (Kwon et al., PagedAttention), 12 September 2023
- Price
- free (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- Apache 2.0
- Official site
- github.com/vllm-project/vllm
- Glossary term
- vLLM
- #13
Jan
Menlo Research, open-source project
- What it is good for
- An open program with a chat window and a Russian interface that works without the internet.
- Why it is on the list
- According to Menlo Research, as of October 2026 Jan has been downloaded more than 7 million times, and the repository has 45 thousand stars on GitHub. Jan, October 2026
- Price
- free (16 October 2026) plans
- Russian language
- yes per the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- jan.ai
In short: models for your hardware
The main limit of a local modellocal LLM. A local LLM is a neural network that runs on your own computer or server, not in a vendor’s cloud, so your data goes nowhere. is memory. The whole model sits in RAM or in the graphics card’s memory, so the choice starts not with a ranking but with how much memory your computer has. The exact calculation for a model and its compression level comes from the memory calculator.
The models’ places come from the open Arena Text ranking, where users compare the answers of two models blindly. The ranking measures answer quality, not hardware needs: Kimi K3 got 18th place in Arena Text on 16 October 2026, while Gemma 4 31B, which fits on a workstation, got 76th place in Arena Text on 16 October 2026. Launch programs have no places and come after the models.
Models up to 16 GB of memory
According to its model card, OpenAI’s gpt-oss-20b runs in 16 GB of memory: it has 21 billion parameters, of which 3.6 billion are active. The Apache 2.0 license allows commercial use.
Gemma 4 came out on 2 April 2026 in four sizes. Google built the small E2B and E4B for phones and other devices: according to the company, they work fully without the internet. Gemma 4 is also under Apache 2.0, while the earlier Gemma 1–3 came under Google’s own terms.
Russian is not officially confirmed by a list for these models: Gemma 4 was trained on more than 140 languages, but Google gives no list. Test the model on your own documents before you choose.
How to choose a model for a MacBook and how much memory to leave for the system is covered in the section “AI on a MacBook”.
Models for a workstation
A workstation or a Mac with a lot of memory handles models of 27–31 billion parameters. This is the level at which a model confidently writes letters, reviews documents, and helps with code inside a company.
Alibaba’s Qwen3.8–27B is published under Apache 2.0. According to its model card, the context is 262,144 tokenstoken. A token is a piece of text that a language model works with: a word, part of a word, or a symbol., expandable to 1 million: a long contract fits into one request whole.
Google describes Gemma 4 31B as a model for personal computers that works without the internet. The Qwen 3 family under Apache 2.0 came out on 29 April 2025, and since then Alibaba has released versions of all sizes under the same license.
Models for a server
According to OpenAI, gpt-oss-120b runs on one graphics card with 80 GB of memory: this is the lower limit of a server model. It has 117 billion parameters, with 5.1 billion active.
Models the size of flagship chats need several servers. DeepSeek released R1 under MIT on 20 January 2025, and on 10 September 2026 it opened DeepSeek-V4.1-Flash with 552 billion parameters. Z.ai published GLM-5.3 with a score of 88.2 on the Terminal Bench 2.1 test, and Moonshot AI opened Kimi K3 with 2.8 trillion parameters and a context of 1 million tokens.
The Alibaba flagship Qwen3.8 with 2.4 trillion parameters, 95 billion of them active, is also open, but under its own license. On 6 October 2026, Mistral AI presented Mistral Large 4 with 1 trillion parameters, and the company promises the weights by the end of October.
Llama 4 from Meta: the Scout and Maverick models have 17 billion active parameters each, and Scout has a context of 10 million tokens. In Arena Text, only an older version of the family has a place, not Llama 4. In Russia, Meta is designated an extremist organization, so a company in Russia should take this into account before choosing the model.
What a server for such a model and load costs is calculated by the local AI cost calculator.
What to run them with: Ollama, LM Studio, llama.cpp, vLLM
Ollama installs and runs a model with one command, MIT license. According to Ollama, as of October 2026 the program is installed more than 9 million times a month. Ollama also has cloud modelscloud AI model. A cloud AI model is a neural network that runs on the servers of a provider. You use it through a website, an app or an API, and you pay for the volume of requests. by subscription: they do not run on your computer, so that is no longer a local launch.
LM Studio gives you a chat window and a model catalog, and its interface is available in Russian. The program itself is closed, while the lms console tool is under MIT. On 16 July 2026, the Bionic agent came out, which works locally or in the cloud on servers in the US: for company data, use only the local mode.
Jan is similar to LM Studio, but it is open under Apache 2.0 and also translated into Russian. According to the developer, as of October 2026 Jan has been downloaded more than 7 million times.
llama.cpp is the base of many such programs and works even on a CPU without a graphics card; the project has 131,000 stars on GitHub. vLLM is needed when a model serves a whole department: in the authors’ paper, throughput grew 2–4 times at the same latencyLLM latency. LLM latency is the delay between a request to a model and its answer: how long you wait for the first word and how long the whole answer takes.. A detailed comparison is on the page about Ollama, LM Studio, and vLLM.
Licenses: what a business may do
Apache 2.0 and MIT allow commercial use, changes, and embeddingembedding. An embedding is a list of numbers by which a computer understands the meaning of a text or a picture and finds similar ones. in your own products; you must keep the license text. gpt-oss, Gemma 4, Qwen3.8–27B, DeepSeek-R1 and DeepSeek-V4.1-Flash, GLM-5.2, and GLM-5.3-Flash are published under them.
The Qwen3.8, GLM-5.3, and Kimi K3 flagships and Llama 4 have their own licenses with conditions. The conditions concern very large services: Kimi K3 needs a separate agreement when the revenue of a service offering the model through an API exceeds $20 million over 12 months, and Llama 4 when the audience exceeds 700 million users a month.
For a Russian company, data law matters more. Under 152-FZ, Russia’s personal data law, since 1 March 2023 an operator must notify Roskomnadzor, the data protection regulator, before a cross-border transfer of personal data, and must store citizens’ data in databases in Russia. A model on your own server in Russia does not transfer data abroad.
Since 1 September 2026, Federal Law 243-FZRussia AI law. Russia’s AI law is Federal Law No. 243-FZ on supporting the development of artificial intelligence technologies, which introduces the concepts of large foundation models, sovereign models and national models. on supporting the development of AI technologies has been in force: the Government will define the cases where sovereign and national models must be used. An overview of open modelsopen weights model. An open weights model is a neural network whose parameter files you can download and run on your own hardware. by family is in the section “Open models”.
How we built this top
- The model’s place in Arena Text on the check date
- How much memory the model needs
- License: can an ordinary company use it with no conditions
- Russian language support and ease of running the program
Order: Arena Text. Places go only to tools whose models are on the board.
Checked:
How we build ratingsQuestions and answers
How much memory does a local model need?
Can an open model be used commercially?
Sources
- Simon Willison, 20 January 2025simonwillison.net
- Simon Willison, 29 April 2025simonwillison.net
- Pravo.ru, ruling of the Tverskoy District Court of Moscow of 21 March 2022 (does not concern WhatsApp or users), 21 March 2022pravo.ru
- Brace Law Firm, 9 July 2023brace-lf.com
- ppt.ru, 26 July 2026ppt.ru
- OpenAI, model card on Hugging Face, 5 August 2025huggingface.co
- Google, company blog, 2 April 2026blog.google
- Qwen, model card on Hugging Face, August 2026huggingface.co
- Arena Text, checked 16 October 2026arena.ai
* Meta is recognized as an extremist organization and its activity is banned in Russia.