Fast enough to serve a whole team from a single server, and in practice often noticeably faster than the cloud: no network detours, dedicated hardware, no shared resources.
Real numbers from the Server S2
A single S2 (2x NVIDIA RTX PRO 6000 Blackwell) runs DeepSeek v4 flash with the full 1M context window and reaches:
- 1,300 tokens/s text generation under parallel requests, 200 tokens/s single stream
- 10,000 tokens/s input processing under parallel requests
- 44 concurrent users in office workloads, plus 10 parallel coding agents
- 60,000 pages per hour of document classification in continuous operation
- 3.4 billion generated tokens per month, plus 26 billion processed input tokens
Preliminary figures as of August 2026, verification is ongoing. All details on the server page.
Our team optimizes for performance
These numbers do not happen by accident: we select models to fit the use case, quantize and configure them for the hardware at hand, and continuously tune the inference stack, KV cache, and batching. How much that matters is shown in our deep dive: Apertus v1.5 70B on a single NVIDIA RTX PRO 6000 Blackwell, with vision, tool calling, and up to 229k tokens of context on just one GPU.
Why on-premise is often faster
The request goes straight from the workstation to the server in your own building: no internet round trip, no queues behind other tenants, no throttling by rate limits. The GPUs work exclusively for your team.
In short
With the right model and proper tuning, a single server serves an entire company, at latencies that beat cloud setups by a clear margin depending on the scenario.
Next steps
- See the servers with all benchmarks and configurations
- Contact us and discuss your performance requirements
Sources and further information:
- Apertus v1.5 70B on an RTX PRO 6000 (performance tuning in detail)
- What does the license include? (model updates and performance)
- On-premise AI for SMEs (hardware and performance)