AI models · data verified October 28, 2026

Best open AI models to run on your own server in 2026

Qwen3.8-27B gives you the most quality on modest hardware: 52 points on the Artificial Analysis index and an Apache 2.0 license. Muse Glimmer 30B and Gemma 4 31B also fit on a single graphics card. The strongest open model, Kimi K3, needs a data centre, and under its license EU companies may not run Llama 4 themselves.

  1. Qwen3.8-27B Alibaba (Qwen) · small

    52 points on the Artificial Analysis index, third among all open models, yet it has only 27 billion parameters and fits on a powerful consumer machine. Apache 2.0 license, image and video input, and Alibaba reports 61.7 on SWE-bench Pro. Thinking is on at a high level by default, so switch it off for simple tasks.

    API price: $0.5 / $3 per 1M tokens (input / output) · Context window: 262K tokens

  2. Muse Glimmer 30B Meta · small

    Apache 2.0 with no EU restrictions and no revenue cap. At 4 bits the language model is under 20 GB and runs on a single graphics card, up to 233 tokens per second on an RTX 5090. Meta reports 76.0% on SWE-Bench Verified, and its knowledge runs to January 2026. Even in Meta's own table, though, it trails Qwen3.6-27B on some tests.

    API price: $0.3 / $1.2 per 1M tokens (input / output) · Context window: 131K tokens

  3. Gemma 4 31B Google · small

    From Google under Apache 2.0. The 4-bit version needs about 17.5 GB of graphics memory, the 8-bit one 34.9 GB. It handles more than 35 languages and scores 84.3% on GPQA Diamond and 80.0% on LiveCodeBench v6. The downsides: knowledge only to January 2025, a 256K-token context and 19.5% on Humanity's Last Exam.

    API price: no published API price per 1M tokens (input / output) · Context window: 256K tokens

  4. Nemotron 3.5 Bolt 74B-A9B NVIDIA · small

    Nemotron 3.5 Bolt from NVIDIA fits into 24 GB of memory thanks to 2-bit packing, and NVIDIA tested it on an RTX 5090. Ollama, LM Studio and llama.cpp support it, and the OpenMDW 1.1 license allows commercial use. Benchmark results so far come only from NVIDIA, and Slovak and Czech are not among its languages.

    API price: no published API price per 1M tokens (input / output) · Context window: 256K tokens

  5. Mistral Medium 3.5 Mistral AI · balanced

    The newest model from Europe’s Mistral with weights already out. It has 128 billion parameters, and Mistral says four GPUs are enough. Mistral reports 77.6% on SWE-bench Verified. The license is a modified MIT, and companies with large revenue fall outside the free terms. The bigger Mistral Large 4 had not released its weights as of 28 October 2026.

    API price: $1.5 / $7.5 per 1M tokens (input / output) · Context window: 256K tokens

  6. Nemotron 3.5 Super VL NVIDIA · balanced

    Nemotron 3.5 Super VL from NVIDIA understands text, images and video and has a 1 million token context. The leaner NVFP4 version runs on one H200 or two H100 cards. It is very fast, Artificial Analysis measured 447 tokens per second. Quality benchmarks so far come only from NVIDIA.

    API price: $0.15 / $0.5 per 1M tokens (input / output) · Context window: 1M tokens

  7. Command A+ Cohere · balanced

    It officially supports 48 languages, including Slovak and Czech, which suits business documents and translation. Apache 2.0 with no revenue cap, and Cohere says that with W4A4 quantization it runs on one B200 or two H100 cards. Its context is only 128K tokens and Cohere has published no benchmark results, so test the quality yourself.

    API price: no published API price per 1M tokens (input / output) · Context window: 128K tokens

  8. DeepSeek V4.1 Flash DeepSeek · fast and cheap

    MIT license, a 1-million-token context, answers up to 384K tokens and image understanding. DeepSeek reports 77.0 on DeepSWE v1.1. On your own server no data goes to China, and the NÚKIB warning does not cover this kind of use. With 552 billion parameters, though, it needs server hardware, not an ordinary PC.

    API price: $0.3 / $1.2 per 1M tokens (input / output) · Context window: 1M tokens

  9. Kimi K3 Moonshot AI · flagship

    The highest Artificial Analysis score of any open model, 60 points in August 2026. It has 2.8 trillion parameters, a 1-million-token context and understands images and video. According to industry press, though, it needs 64 or more accelerators, so a small company practically cannot run it. The license is not plain MIT: large service providers need a separate agreement.

    API price: $3 / $15 per 1M tokens (input / output) · Context window: 1M tokens

  10. Beam 1.0 Reflection AI · flagship

    Beam 1.0 from Reflection AI has a permissive Apache 2.0 license, a 1 million token context and, according to the vendor, strong coding results such as 81.6% on SWE-bench Verified. It is a data center giant, though, as even the smallest version is about 280 GB. It understands text only, and independent tests are not out yet.

    API price: $0.25 / $1 per 1M tokens (input / output) · Context window: 1M tokens

  11. gpt-oss-120b OpenAI · reasoning

    It runs on a single 80 GB card such as an H100 and uses the Apache 2.0 license. You set thinking to low, medium or high and see the full chain of reasoning. It is older, though: knowledge to June 2024, text only, a 131K-token context, and trained mostly on English, so smaller languages may suffer. The smaller gpt-oss-20b needs only 16 GB of memory.

    API price: no published API price per 1M tokens (input / output) · Context window: 131K tokens

Llama 4 and EU companies

The Llama 4 license grants no rights to its multimodal models to people living in the EU or companies headquartered there. So a European company may not run, fine-tune or build a product on Llama 4 Maverick or Scout itself. End users of finished apps are not affected. Meta hasn’t released a newer Llama since. The replacements are Muse Glimmer 30B under Apache 2.0 and Muse Spark Open, whose license does not exclude the EU.

What hardware you need

  • One graphics card: Gemma 4 31B (about 17.5 GB at 4 bits), Muse Glimmer 30B (under 20 GB), Gemma 4 26B A4B (about 14.4 GB) and gpt-oss-20b (16 GB).
  • One 80 GB server card: gpt-oss-120b.
  • Several server GPUs: Mistral Medium 3.5 (four GPUs), Command A+ (one B200 or two H100) and Mistral Small 4, which despite its name has 119 billion parameters.
  • A data centre: Muse Spark Open (about 537 GB even compressed), DeepSeek V4 Pro (a node with four GB300 chips), Kimi K3 and Qwen3.8-2.4T-A95B.

Licenses: what to watch out for

  • Apache 2.0 and MIT are the most permissive, allowing commercial use and changes. Qwen3.8-27B, Gemma 4, Muse Glimmer 30B, Command A+, Mistral Small 4, gpt-oss and the DeepSeek models use them.
  • Custom licenses with conditions: Kimi K3 (very large products must display the Kimi K3 name), Muse Spark Open (a “Built with Muse Spark Open” label and users 18 and over), Mistral Medium 3.5 (a revenue cap) and Qwen3.8-2.4T-A95B (conditions tied to company size).
  • For GLM-5.3 the license is unclear: Hugging Face lists a custom license, Mistral's docs say MIT. Read the terms before commercial use.

How we chose

We measure quality with the independent Artificial Analysis index and vendor tests, but whether a small or mid-sized company can actually run the model counts just as much. So models that run on one graphics card or a few server GPUs come first, and data-centre giants sit lower even with better scores. We also weigh the license (commercial use, EU restrictions, revenue caps) and support for smaller languages such as Slovak and Czech. Llama 4, because of its license, and Mistral Large 4, with no published weights, are not ranked.

FAQ

Which open AI model is the best?

On the Artificial Analysis index, Kimi K3 with 60 points in August 2026, but it needs a data centre. On ordinary company hardware the best is Qwen3.8-27B with 52 points, which fits on a powerful computer.

Can I use an open model commercially?

With Apache 2.0 and MIT, yes, with no fees and no revenue cap. Custom licenses such as those of Kimi K3, Muse Spark Open or Mistral Medium 3.5 come with conditions, such as product labelling or a revenue cap. Companies headquartered in the EU may not run, fine-tune or build products on Llama 4 themselves.

Is it safe to run a Chinese model such as DeepSeek or Qwen on my own server?

As far as data goes, yes: the model runs on your side and sends nothing to the vendor. Czechia's NÚKIB explicitly said its DeepSeek warning does not cover open models run locally with no connection to DeepSeek servers. The DeepSeek app and API are a different case, as they store data in China.

Sources

Prices and models change often. The data comes from official vendor pricing pages and documentation, and every page shows when it was last checked. Before a larger purchase, confirm the price with the vendor.

← All AI models

Not sure which model fits your business?

I can help you pick a model and plan based on what you actually need, set it up safely for company data and train your team to use it well. Let’s start with an intro consultation.

See pricing jozef@walterstein.eu