Home/Installed models
What runs on your premises. Not what tops a leaderboard.
Open models with downloadable weights, installed on the machine and updated quarterly. Every row was checked against the model card at its maker.
- Collected
- 29.08.2026
- Models
- 5
- Refresh
- quarterly
01 — THE CATALOGUE
Open, installable, yours.
An open model stays on the machine after the service contract ends: that is what separates a downloaded weight from an API key.
Proof| Model | Size | GPU memory | Context | Licence | Machines |
|---|---|---|---|---|---|
| Qwen 3.8 27BAlibaba · flagship | 27 B · dense | ≈ 56 Go (BF16) | 256 k tokens | Apache 2.0 | N1 · N2 · N4 |
| GLM-5.3-FlashZ.ai | 320 B · 18 B actifs | ≈ 320 Go (BF16) | 256 k tokens | MIT | N2 · N4 |
| DeepSeek V4 FlashDeepSeek | 284 B · 13 B actifs | ≈ 291 Go (BF16) | 1 M tokens | MIT | N4 |
| Apertus 70BEPFL / ETH / CSCS · Suisse | 71 B · dense | ≈ 142 Go (BF16) | 131 k tokens | Apache 2.0 | N2 · N4 |
| Apertus v1.5 8BEPFL / ETH / CSCS · Suisse | 9 B · dense | ≈ 18 Go (BF16) | 131 k tokens | Apache 2.0 | N1 · N2 · N4 |
02 — WHAT WE DO NOT INSTALL
The models that top the leaderboards do not fit in a building.
The “max” versions that top the leaderboards (Kimi K3: 2.8 trillion parameters, Qwen 3.8 Max, GLM-5.3) need clusters of several dozen cards and remain API services. They serve here as a level reference. What the machine installs is the catalogue above — the most powerful models an organisation can actually run inside its own walls.
Sources for the rankings quoted, collected on 29.08.2026 : LMArena, Artificial Analysis.
Gap to closed models, cost per task, indices: all on the Proof page, with source and date.