Buy Kosmik-built GPU servers configured for local LLM inference. From a single RTX 5090 workstation to an 8-GPU RTX PRO 6000 EPYC node.
Used hardware, one-year warranty, and units in stock. Tell us which configuration fits your workload and we will reserve it.
01 The lineup
Every machine below lists hardware first, then performance measured on its reference model — so context, tok/s, throughput, and power draw apply only to that model.
190,000 Kč Price excl. VAT
In stock : 1 unit
Performance on reference model
Context, speed, throughput, and power below apply only to this model.
350,000 Kč Price excl. VAT
In stock : 6 units
Performance on reference model
Context, speed, throughput, and power below apply only to this model.
670,000 Kč Price excl. VAT
In stock : 4 units
Performance on reference model
Context, speed, throughput, and power below apply only to this model.
1,390,000 Kč Price excl. VAT
In stock : 1 unit
Performance on reference model
Context, speed, throughput, and power below apply only to this model.
2,900,000 Kč Price excl. VAT
In stock : 1 unit
Performance on reference model
Context, speed, throughput, and power below apply only to this model.
02 Compare
Hardware rows describe the machine. Performance rows (from reference model downward) apply only to that reference model.
| Comet | Pulsar | Binary | Quasar | Supernova | |
|---|---|---|---|---|---|
| Hardware | |||||
| Motherboard | AM5, PCIe 5.0 ×16 | AM5, PCIe 5.0 ×16 | AM5, 2× PCIe 5.0 ×8 | Server-grade | Server-grade |
| CPU | Ryzen 7900 (24 threads) | Ryzen 7900 (24 threads) | Ryzen 7900 (24 threads) | 2× AMD EPYC (96 threads total) | 2× AMD EPYC (96 threads total) |
| VRAM | 32 GB (1× RTX 5090) | 96 GB (1× RTX 6000 Pro Max-Q) | 192 GB (2× RTX 6000 Pro Max-Q) | 384 GB (4× RTX 6000 Pro Max-Q) | 768 GB (8× RTX 6000 Pro Max-Q) |
| SSD | 2 TB NVMe | 4 TB NVMe | 4 TB NVMe | 4 TB NVMe | 8 TB NVMe |
| RAM | 128 GB DDR5 RAM | 128 GB DDR5 RAM | 128 GB DDR5 RAM | 256 GB DDR4 ECC RDIMM | 1 TB DDR4 ECC RDIMM |
| Performance on reference model | |||||
| Reference model | Qwen 3.8 27B NVFP4 | Qwen3.8-Flash-Next NVFP4 + MTP | GLM-5.3-Flash NVFP4 | GLM-5.3-Flash NVFP4 | GLM-5.3 full NVFP4 |
| Max context | 160k | 262k | 1M | 1M | 1M |
| Single tok/s | 120–140 tok/s | 150–170 tok/s | 170–190 tok/s | 170–190 tok/s | 120–130 tok/s |
| Max throughput | 550–620 tok/s (4c) | 550–620 tok/s (8c) | 900–1000 tok/s (16c) | 1800–2000 tok/s (32c) | 1000–1150 tok/s (128c) |
| Optimized power (AI load) | 540 W | 370 W | 650 W | 1.3 kW | 2.7 kW |
| Price excl. VAT | 190,000 Kč | 350,000 Kč | 670,000 Kč | 1,390,000 Kč | 2,900,000 Kč |
| In stock | 1 unit | 6 units | 4 units | 1 unit | 1 unit |
Important – All listed servers are used. Warranty: one year.
Email info@koscompute.com with the model you want. We reply with availability and next steps.
