Comparisons
Open-source LLMs compared
Five open-weight LLM families, scored uniformly on the same seven criteria - with a source link and date for every entry, so you can verify the numbers yourself. This category is aimed at developers and technical decision-makers building an application on top of a model - not end users of a finished chat app (see our "AI Chat Assistants" category for that). Switch between the individual and business perspective at the top of the table. At the bottom you'll find scenario recommendations instead of a single "winner": which model fits depends on your specific situation.
A comparison of open-source LLMs weighs model families like Llama, Mistral, DeepSeek, Qwen and Gemma against uniform criteria - from license freedom to hardware requirements to ecosystem support - for developers and technical decision-makers building their own applications on a model.
How we score →Last data review: 07/05/2026, 06:00 PM
| Tool | Price | Value for money | License & usage freedom | Performance & context window | Hardware requirements & hosting effort | Ecosystem & tooling | Multimodality | Community & enterprise support | Ideal for | Details |
|---|---|---|---|---|---|---|---|---|---|---|
Llama Meta | Weights free · API hosting ~$0.08-2.19/M tokens depending on provider/model size | 3 of 5 | 2 of 5 | 3 of 5 | 2 of 5 | 3 of 5 | 2 of 5 | 4 of 5 | Anyone looking for an already-established, broadly hosted model for a one-off project +1 | |
Mistral Mistral AI | Weights free · La Plateforme API from ~$2/M input tokens (Large 3) | 4 of 5 | 3 of 5 | 4 of 5 | 4 of 5 | 5 of 5 | 3 of 5 | 5 of 5 | Anyone wanting to experiment on ordinary hardware +1 | |
DeepSeek DeepSeek | Weights free (MIT) · Own API from $0.14/M input tokens (V4-Flash) | 5 of 5 | 5 of 5 | 4 of 5 | 3 of 5 | 3 of 5 | 2 of 5 | 3 of 5 | Price-conscious developers with their own hosting +1 | |
Qwen Alibaba | Weights free (Apache 2.0) · DashScope API with a time-limited free quota | 4 of 5 | 5 of 5 | 3 of 5 | 4 of 5 | 5 of 5 | 5 of 5 | 3 of 5 | Anyone wanting the most versatile, unrestricted-license model +1 | |
Gemma | Weights free (Apache 2.0 as of Gemma 4) · API from ~$0.14/M input tokens | 4 of 5 | 4 of 5 | 4 of 5 | 5 of 5 | 5 of 5 | 4 of 5 | 4 of 5 | Newcomers who want to experiment on ordinary or even mobile hardware +1 | |
Which tool fits you?
Maximum cost efficiency for self-hosting
DeepSeek
The cheapest API pricing and a fully permissive MIT license in this comparison.
Multimodal applications (image, audio, video)
Qwen
The broadest multimodality of all five model families, including real-time voice output.
Needs to run on weak or mobile hardware
Gemma
The smallest model size runs quantized at roughly 1GB.
Key takeaways
- Mistral is the top choice for strict EU data sovereignty - it runs its own data centers in France and partners with European cloud providers (Scaleway/OVHcloud).
- DeepSeek offers the cheapest API pricing and a fully permissive MIT license - the most cost-efficient option in this comparison.
- Gemma already runs quantized at around 1GB, making it suitable even for weak or mobile hardware.
- Qwen offers the broadest multimodality (image, audio, video, real-time voice) among the five model families.
- This category targets developers building their own applications - if you're looking for a ready-made chat app, see our "AI Chat Assistants" category instead.
→ Have a custom application built on top of it: Custom Applications
Frequently asked questions
What's the difference between open-weight and closed-source LLMs?
Open-weight LLMs like Llama, Mistral, or DeepSeek publish their model weights for download - you can host and customize them yourself. Closed-source LLMs like GPT or Claude are only accessible via the vendor's API, with no access to the model weights themselves.
Which open-source LLM is best for EU data sovereignty?
Mistral AI runs its own data centers in France and partners with European cloud providers like Scaleway and OVHcloud - the strongest option in this comparison for organizations with strict EU data-sovereignty requirements.
Which open-source LLM is the cheapest?
DeepSeek offers the cheapest API pricing among the five model families compared, combined with a fully permissive MIT license with no usage restrictions.
Do I need an expensive GPU to self-host an open LLM?
It depends on the model and size. Gemma, the smallest model family in this comparison, already runs quantized on roughly 1GB of VRAM, while larger Llama or Mistral variants need considerably more hardware - see the "hardware requirements" criterion column in the table.