AI ranking
Local / open-source AI
Open or locally runnable models (open weights, open source, local runtimes). Compared on the same quality criteria as the general ranking.
Top visuel
- 01 Qwen3.8 27B 84.97
- 02 Qwen3.5 27B 81.80
- 03 Qwen3.5 35B A3b 76.71
- 04 Qwen3 32b 75.76
- 05 Qwen3 30B A3b Thinking 2507 75.04
- 06 Qwen3.5 9B 70.66
- 07 Qwen3.6 27B 66.58
- 08 Llama 4 Maverick 17B 128e Instruct 65.38
- 09 Qwen3 30B A3b 64.23
- 10 Mistral Small 3.1 24B Instruct 2503 61.19
Models beyond roughly 35 billion parameters are not ranked in the local categories. Le Recul favours models that can genuinely be run locally, without heavy professional infrastructure.
Compare models. Cochez jusqu'à 3 modèles dans le tableau ci-dessous, puis lancez la comparaison : qui gagne sur quel usage, de combien, et avec quelle solidité.
| # | Model | Vendor | Score | Assessment | |
|---|---|---|---|---|---|
| 01 | Alibaba | 84.97 / 100 | Partial | ||
| 02 | Alibaba | 81.80 / 100 | Reliable | ||
| 03 | Alibaba | 76.71 / 100 | Reliable | ||
| 04 | Alibaba | 75.76 / 100 | Reliable | ||
| 05 | Alibaba | 75.04 / 100 | Reliable | ||
| 06 | Alibaba | 70.66 / 100 | Reliable | ||
| 07 | Alibaba | 66.58 / 100 | Reliable | ||
| 08 | Meta | 65.38 / 100 | Partial | ||
| 09 | Alibaba | 64.23 / 100 | Reliable | ||
| 10 | Mistral | 61.19 / 100 | Partial | ||
| 11 | Alibaba | 59.05 / 100 | Partial | ||
| 12 | Olmo 3.1 32b instruct | Allen institute for ai | 57.90 / 100 | Reliable | |
| 13 | 57.17 / 100 | Partial | |||
| 14 | Mistral | 56.12 / 100 | Partial | ||
| 15 | Meta | 55.69 / 100 | Reliable | ||
| 16 | Alibaba | 53.10 / 100 | Partial | ||
| 17 | Alibaba | 52.91 / 100 | Partial | ||
| 18 | Olmo 2 0325 32B Instruct | Allenai | 52.27 / 100 | Partial | |
| 19 | Alibaba | 49.40 / 100 | Partial | ||
| 20 | Alibaba | 49.09 / 100 | Partial | ||
| 21 | Alibaba | 46.89 / 100 | Partial | ||
| 22 | Alibaba | 46.78 / 100 | Partial | ||
| 23 | Alibaba | 46.73 / 100 | Partial | ||
| 24 | Allenai | 46.68 / 100 | Partial | ||
| 25 | Alibaba | 46.63 / 100 | Partial | ||
| 26 | 46.53 / 100 | Reliable | |||
| 27 | Alibaba | 46.46 / 100 | Reliable | ||
| 28 | Alibaba | 45.00 / 100 | Partial | ||
| 29 | Alibaba | 44.64 / 100 | Reliable | ||
| 30 | 42.52 / 100 | Reliable | |||
| 31 | Olmo 3.1 32b think | Allen institute for ai | 42.50 / 100 | Reliable | |
| 32 | 01.ai | 42.32 / 100 | Reliable | ||
| 33 | Alibaba | 42.28 / 100 | Partial | ||
| 34 | Olmo 3 32b think | Allen institute for ai | 42.27 / 100 | Reliable | |
| 35 | Alibaba | 42.24 / 100 | Reliable | ||
| 36 | 42.05 / 100 | Reliable | |||
| 37 | Internlm2 5 20B Chat | Shanghai ai lab | 41.61 / 100 | Partial | |
| 38 | Alibaba | 40.44 / 100 | Reliable | ||
| 39 | 40.43 / 100 | Partial | |||
| 40 | 40.27 / 100 | Partial | |||
| 41 | 40.21 / 100 | Partial | |||
| 42 | Alibaba | 40.17 / 100 | Partial | ||
| 43 | Alibaba | 39.01 / 100 | Reliable | ||
| 44 | Alibaba | 38.38 / 100 | Partial | ||
| 45 | Alibaba | 38.17 / 100 | Partial | ||
| 46 | Alibaba | 37.87 / 100 | Partial | ||
| 47 | NVIDIA | 37.82 / 100 | Partial | ||
| 48 | 37.62 / 100 | Reliable | |||
| 49 | Alibaba | 37.14 / 100 | Reliable | ||
| 50 | Alibaba | 36.48 / 100 | Reliable |
Score indicatif, pondéré à partir de sources publiques. Méthodologie transparente. Voir la page Classement IA pour la méthodologie complète et les autres catégories.
Open-source and local AI ranking
Le Recul's Open source ranking gathers the models that are open or can be run locally: open weights, permissive licences, local runtimes. They are compared on the same general-purpose grid as closed models — overall quality, reasoning, knowledge, multimodal, code, cost, factual accuracy, speed, stability, accessibility — so that no indulgent scale is applied to them.
The value of an open model does not lie in its score alone. It lies in what you can do with it: run it on your own machines, keep your data, avoid depending on a vendor, or simply carry on using it if the vendor changes its terms. Those advantages do not show up in a ranking, but they often weigh more than a few points of difference.
In return, « open » means neither free nor effortless. It takes hardware, configuration time, and accepting a performance gap that is narrowing but persists on the most demanding tasks. The score places the model; the decision to use it depends on your constraints.
What this ranking is for
This ranking helps you spot the open models that are genuinely competitive, and separate those that hold up in real use from those that only shine on a launch-day benchmark.
It also helps frame a sovereignty or confidentiality decision: if your data must not leave, the question is not « which is the best model » but « which is the best model I can host ».
Comment lire le classement
- Overall quality and reasoning, measured on the same basis as closed models.
- Knowledge, multimodal and code, depending on public source coverage.
- Real cost of running it, hardware included — not just the absence of a licence fee.
- Stability and accessibility: availability of weights, formats, runtimes.
- Evaluation status (Reliable / Partial / Insufficient) depending on coverage.
FAQ — Local / open-source AI
Are open-source AI models as good as closed ones?
The gap has narrowed sharply on everyday uses and remains clearer on long or complex tasks. The ranking shows it without inflating or downplaying it, since the grid is identical.
Are « open source » and « open weights » the same thing?
No. Many so-called open models publish their weights under a restrictive licence, without the training data or code. That is useful for running them locally, but it is not open source in the strict sense.
Do you need an expensive GPU to run them?
Not always. Compact models run on a consumer GPU or a recent Apple Silicon Mac. The largest ones demand significant hardware, which is part of the real cost.
Is a local model safer for my data?
In principle yes, since nothing leaves your machine. That shifts the responsibility onto your own security: updates, access, backups.
How often is the ranking updated?
About every 25 hours, automatically. The update date appears at the top of the page.