Sizes
The standard configurations, from the smallest to the largest. The customer buys the machine: we install it and keep it updated.
Standard configurations
Performance figures are theoretical estimates from technical data: they are measured on the first real machine before being promised.
| Size | Who it is for | Hardware | Main model | Indicative users |
|---|---|---|---|---|
| XS Studio | Professional firms, small departments | ASUS Ascent GX10 or DGX Spark, 128 GB unified | Qwen3.6-35B-A3B FP8, 0.6B embedding and reranker | up to 10-15 |
| XSx2 Studio Pro | Firms and small companies that want the best model at a low price | Two DGX Spark or GX10 linked with a direct 200 Gb cable | Qwen3.8-Flash-Next NVFP4 (experimental), 0.6B embedding and reranker | up to 60-80 |
| S Department | One department of a small company, pilot | Dell Pro Max T2, Core Ultra 7, 64 GB RAM, 2 TB SSD, RTX PRO 5000 48 GB | Qwen3.8-27B FP8, 0.6B embedding and reranker | up to 40-60 |
| M Company | Companies from 50 to 200 people, several departments | Dell Pro Max T2, Core Ultra 9, 128 GB RAM, 2 TB SSD, RTX PRO 6000 96 GB | Qwen3.8-27B FP8 with plenty of memory for users, or Qwen3.5-122B-A10B in 4 bit for more quality; 4B embedding and reranker | up to 150-200 |
| L Group | Groups with several sites, intensive use | Thomas-Krenn 2U server, EPYC 9355, 256 GB RAM, two RTX PRO 6000 Server | Qwen3.5-122B-A10B FP8; 4B or 8B embedding and reranker | up to 400-500 |
| XL Frontier | On request only | Thomas-Krenn 4U server, four RTX PRO 6000 Server, 384 GB RAM | GLM-5.3-Flash; 8B embedding and reranker | hundreds |
- XS only takes MoE models with few active parameters: memory bandwidth is 273 GB/s, and with few concurrent users it must be verified with a load test.
- XSx2 is off list: it needs SGLang with two patches and an experimental model. It comes back when it runs on vLLM and the licence is settled.
- Workstation RAM is enough at 64 GB for size S and 128 GB for M. ECC memory costs almost three times as much and is used in servers only.
Which one to choose
- Pilot: almost always size S. If the customer already has a server with free slots and enough power, the RTX PRO 5000 card alone is enough.
- Firms with few users and a small budget: XS. Its only limit: MoE models with few active parameters and few concurrent users.
- Groups watching the budget: start with the pilot on S and move to M only after the pilot has shown its value.
- L and XL are for intensive use and several sites; a model change is a scheduled operation, with weights downloaded first and a load test afterwards.
How it is bought
The customer buys a complete system from a single vendor (Dell, Lenovo or Thomas-Krenn) or through a local integrator. Loose GPUs only for servers the customer already has and the integrator confirms as compatible. We do not resell hardware: we install, measure and maintain.
The customer pays for the hardware to its own vendor: we do not publish list prices here, because they change with configuration and delivery times.
The service
The borc licence is quoted by size and seats. The Pricing page explains what the contract includes.