Free tool
Which open-weight model can you actually use?
Benchmark sites tell you which model is smartest. This table answers the question that matters in a regulated European organisation: which one may you deploy — under what license, on what hardware, in which languages. Deployability data, curated and dated; for capability benchmarks we link the people who measure them.
Show only models that…
Showing 18 of 18 models
| Model | Tier | License | On-prem hardware | Managed EU API | FR/NL | Best for |
|---|---|---|---|---|---|---|
Mistral Small 3.2 (24B) Mistral AI (FR) European vendor; strong all-rounder for its size. | Strong | Permissive Apache 2.0 | Single 24–48 GB GPU | EU-native (Mistral) | Good | Assistant / chat Document Q&A (RAG) Data extraction |
Magistral Small (24B) Mistral AI (FR) Reasoning variant of Mistral Small. | Strong | Permissive Apache 2.0 | Single 24–48 GB GPU | EU-native (Mistral) | Good | Reasoning Assistant / chat |
Devstral Small (24B) Mistral AI (FR) Tuned for software-engineering agents. | Strong | Permissive Apache 2.0 | Single 24–48 GB GPU | EU-native (Mistral) | Usable | Code |
Mistral NeMo (12B) Mistral AI (FR) / NVIDIA Small enough for modest hardware; multilingual. | Compact | Permissive Apache 2.0 | Laptop / small GPU (≤16 GB) | EU-native (Mistral) | Good | Assistant / chat Data extraction |
EuroLLM 9B EuroLLM consortium (EU) EU-funded, built for all official EU languages — including Dutch. | Compact | Permissive Apache 2.0 | Laptop / small GPU (≤16 GB) | Check availability | Good | Assistant / chat |
Llama 3.3 70B Meta License conditions are light for mid-market use (scale threshold, naming). | Strong | Conditions Llama Community License | 1× 80 GB or 2× 48 GB | Common on EU clouds | Usable | Assistant / chat Document Q&A (RAG) |
Llama 4 Scout (109B MoE) Meta Multimodal input, very long context; MoE keeps serving cost down. | Strong | Conditions Llama Community License | 1× 80 GB or 2× 48 GB | Common on EU clouds | Usable | Assistant / chat Document Q&A (RAG) |
Qwen3 32B Alibaba Strong multilingual coverage. | Strong | Permissive Apache 2.0 | Single 24–48 GB GPU | Common on EU clouds | Good | Assistant / chat Document Q&A (RAG) Data extraction |
Qwen3 235B-A22B (MoE) Alibaba Frontier-class open weights; serious hardware. | Frontier | Permissive Apache 2.0 | Multi-GPU server | Common on EU clouds | Good | Assistant / chat Code Reasoning |
DeepSeek V3.1 (MoE) DeepSeek Weights are MIT; self-hosting avoids the questions the hosted API raises. | Frontier | Permissive MIT | Multi-GPU server | Common on EU clouds | Usable | Assistant / chat Code |
DeepSeek R1 (0528) DeepSeek Open reasoning model; distilled variants exist for smaller hardware. | Frontier | Permissive MIT | Multi-GPU server | Common on EU clouds | Usable | Reasoning Code |
gpt-oss-120b (MoE) OpenAI Runs on a single 80 GB GPU despite its size (MoE). | Strong | Permissive Apache 2.0 | 1× 80 GB or 2× 48 GB | Check availability | Usable | Reasoning Assistant / chat Code |
gpt-oss-20b OpenAI Reasoning at laptop scale. | Compact | Permissive Apache 2.0 | Laptop / small GPU (≤16 GB) | Check availability | Usable | Assistant / chat Code |
Gemma 3 27B Multimodal input; use-policy conditions attached. | Strong | Conditions Gemma Terms of Use | Single 24–48 GB GPU | Check availability | Usable | Assistant / chat Document Q&A (RAG) |
Phi-4 (14B) Microsoft Punches above its size in English; weaker FR/NL. | Compact | Permissive MIT | Laptop / small GPU (≤16 GB) | Check availability | Weak | Reasoning Data extraction Code |
Granite 3.3 8B IBM Enterprise-oriented; conservative and documented. | Compact | Permissive Apache 2.0 | Laptop / small GPU (≤16 GB) | Check availability | Usable | Document Q&A (RAG) Data extraction |
GLM-4.5-Air (106B MoE) Z.ai (Zhipu) Agent-oriented open MoE. | Strong | Permissive MIT | 1× 80 GB or 2× 48 GB | Check availability | Usable | Code Reasoning |
Kimi K2 (1T MoE) Moonshot AI Attribution condition kicks in only at very large scale. | Frontier | Permissive Modified MIT | Multi-GPU server | Check availability | Usable | Code Assistant / chat |
How to read this table
- · Hardware tiers assume 4-bit quantised inference and are indicative — sizing for production is a conversation, not a table row.
- · Every model here has downloadable weights, so all of them can run on your own infrastructure; the "Managed EU API" column is for when you want someone else to run it inside the EU.
- · "Conditions" licenses (Llama, Gemma) are workable for most mid-market organisations — but have legal read the terms before you build on them.
- · We publish no capability scores on purpose: they change weekly and others measure them well. See Artificial Analysis for current benchmarks.
Data last reviewed: 28 August 2026. Reviewed monthly — licenses and availability change.
Shortlist made? The next question is the pipeline.
A model is a component. KnowledgeRAG turns one into a document assistant that respects your access rights and stays inside your walls — and our sovereign AI service builds the platform around it.
Mohamed Ben Haddou
Founder & CEO · Independent AI Expert for the European Commission
You talk to Mohamed, not a sales team — and the person who scopes your work is the one who delivers it.