Most rankings of the best open-source LLM answer a question almost nobody is asking: what scores highest? The answer is Chinese, it has a trillion parameters, and you cannot run it on anything you own.
The question worth answering is which model you can legally deploy and physically load. That depends on the job. Here is the short version.
| Best on capability | MiMo-V2.6-Pro, from Xiaomi in China. 46 on the Artificial Analysis index, MIT licence. |
| Best you can actually sign | Mistral Small 4. Apache 2.0, 119B total and 6.5B active, 256k context. |
| Best on one machine | Devstral Small 2 24B for code, Ministral 3 14B for everything else. |
| Do not pick | Teuken-7B, non-commercial. Five more ship under a US company’s licence. |
| Read this first | EU companies cannot license Llama 4’s multimodal models. Meta’s newer Muse Glimmer is Apache 2.0. |
Which Open-Source LLM Scores Highest?
On Artificial Analysis’ open-weights index, now version 4.3.2, Xiaomi’s MiMo-V2.6-Pro leads at 46, ahead of Z.ai’s GLM-5.3 at 45 and Moonshot’s Kimi K3 at 44. The best from outside China is South Korea’s Motif 3, at an estimated 34. Europe’s best is Mistral Medium 3.5, at 14, and it is not Apache 2.0.
Mistral Large 4 is not in that count. It scores 38 as an API preview, but its weights are not public yet; Mistral says they come by the end of October. The leader’s MIT licence lets anyone use it, but a trillion parameters will not fit on anything an ordinary company owns. Further down the list, the licence is the trap.
The Best Open-Source LLM for Each Job
There is no single best model, only a best one for each job. The licence column decides more of these rows than the benchmark column does.
| If you are doing this | Use this | Licence | Why |
|---|---|---|---|
| General work | Mistral Small 4 (119B, 6.5B active) | Apache 2.0 | Europe’s best mix of permissive terms, current training and a 256k context |
| Coding | Devstral Small 2 24B | Apache 2.0 | 68.0% on SWE-Bench Verified, on hardware a small team already owns |
| All 24 EU languages | EuroLLM-22B-Instruct-2512 | Apache 2.0 | 35 languages. The 32,768-token context is the catch |
| Your own machine | Ministral 3 14B | Apache 2.0 | Best English score in our 16GB MacBook Air test, 94 of 100 |
| Long German or English documents | Kolibri-1 | Apache 2.0 | Aleph Alpha, 2 October. A million-token context, but it needs an H200 or two H100s |
| One European language | Bielik 11B v3.0 | Apache 2.0 | Polish-optimised, and its 4-bit build fits one consumer graphics card |
| Translation | salamandraTA-7b-instruct | GPL-3.0 | 40 languages. Commercial use allowed, with copyleft obligations |
| Content moderation | Shieldstral 1.0 3B | Apache 2.0 | 88.1% F1 on WildGuardTest, and it checks images as well as text |
| Maximum capability | Mistral Medium 3.5 | Modified MIT | 77.6% on SWE-Bench Verified, but no rights once your company passes $20m monthly revenue |
The Licence Trap That Will Cost You
Open weights means you can download the parameters. Open source means you can also use them commercially without asking, under a licence an ordinary lawyer recognises. Several models sold as European sovereign AI are the first and not the second.
Meta’s Llama 4 Acceptable Use Policy ends with this: “the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union.” That covers Llama 4’s multimodal models. Meta’s newer Muse Glimmer, a 30B multimodal model, is Apache 2.0 with no such clause.
Europe’s own public models let California set the terms too. Bulgaria’s BgGPT and Ukraine’s MamayLM ship under Google’s Gemma terms; Greece’s Krikri and Poland’s Llama-PLLuM under Meta’s Llama 3.1 licence, Finland’s Poro 2 under Llama 3.3. All five are publicly funded and marketed as sovereign capability, with terms set in Mountain View and Menlo Park.
And the trap for careful readers: Ministral-8B-Instruct-2410 is not open source, because it ships under the Mistral Research License, while Ministral-3-3B-Instruct-2512 is Apache 2.0. Both pulled more than 100,000 downloads last month. Check the licence on the exact repository, never the family.
The Model That Will Not Run in Your Tool
Apertus v1.5, from Switzerland, offers what almost nobody else does: published training data, plus image and audio input and a 262,144-token context. It went generally available on 17 September and pulled 698,867 downloads in 30 days.
But llama.cpp, Ollama and LM Studio still cannot load it. The model card supports Transformers and vLLM, through Swiss AI’s own patched builds, and the four community GGUF builds all strip out the vision and audio. Pick Apertus for the licence and the transparency, not for 1.5 features in the tool you already have open. Our Apertus test has the detail.
Why the Best Model Is Never the Most Used
Europe’s largest open model, Mistral Large 3 at 675 billion parameters, pulled 13,014 downloads last month across its four official weight repositories. Mistral’s 7B Instruct v0.3, from May 2024, pulled 2.06 million. That is 158 to one, a hardware judgement rather than a quality one.
The better signal is which builds people shrink. Of Large 3’s downloads, 83% were Nvidia’s NVFP4 quantisation, and JetBrains’ own 8-bit build of its Mellum2 coding model outdraws the original ten to one. Nobody quantises a model they are not putting into production. Every European model and who funds it is in our European LLM map.
How to Pick, in Three Steps
- Start with the licence, not the benchmark. Decide what your organisation can sign. That removes more options than any score will.
- Then check what fits on the hardware you own. The 158-to-one gap above is the market telling you this decides adoption.
- Only then look at scores. And check the exact repository, not the model family, before you deploy.
Every open model above 34 on that index is Chinese, and Mistral Large 4 may change that at the end of October. What Europe has today is licences you can sign, training data you can inspect and models sized for ordinary companies. For most European businesses that is worth more than thirty points of index score.
We test these models one at a time, most recently on a 16GB MacBook Air.
Frequently Asked Questions
On capability, Xiaomi’s MiMo-V2.6-Pro, at 46 on the Artificial Analysis v4.3.2 index, ahead of Z.ai’s GLM-5.3 and Moonshot’s Kimi K3. It is MIT-licensed but needs data-centre hardware. For a European company, Mistral Small 4 under Apache 2.0. The top of the field is Chinese; Europe’s best published weights score 14.
Devstral Small 2 24B for code and Ministral 3 14B for general use, both Apache 2.0. Ministral 3 14B scored 94 out of 100 on English tasks in our 16GB MacBook Air test. Apertus 1.5’s image and audio features still do not work in Ollama or LM Studio.
Under OSI-approved licences with commercial use permitted: Mistral Large 3, Mistral Small 4, Devstral Small 2, Ministral 3, Shieldstral, Apertus, EuroLLM, Kolibri-1, Bielik, ALIA, Salamandra and Velvet. Not open source: Teuken-7B v0.6 and Ministral-8B are non-commercial, BgGPT and MamayLM use Google’s Gemma terms, Krikri and Poro 2 Meta’s Llama licences, and Mistral Medium 3.5 withdraws its rights above $20m monthly revenue.
Only with data-centre hardware. It is 675 billion parameters, 41 billion active, under Apache 2.0, and its official weights pulled 13,014 downloads in 30 days against 2.06 million for Mistral’s 7B. Of those, 83% were an Nvidia NVFP4 quantisation. Its successor, Mistral Large 4, is due by the end of October.
Author: Akos Szima
See Also:
European LLMs in 2026: The Complete Map
Chinese LLMs in 2026: The Complete Map
