At least 3000 W usable power · server airflow or factory liquid cooling
Model placement
For DeepSeek: four GPUs in one group, with part of the model in RAM
We test speed and your workloads before handover. Shipping, import and complex integrations are charged separately.
What is included in the price and how is it calculated?
Estimate dated 12 October 2026: the complete hardware system plus allowances for assembly, initial setup of the selected model and testing. Handover in China. Shipping, import, ongoing support and complex integrations are calculated separately.
This preliminary range uses public component and base-platform prices, budget allowances and a 10–15% contingency. Before ordering, the quotation will confirm final price, compatibility, availability, lead time and warranty. This range is not a system integrator’s quotation.
Factory PX power system, variant matched to the selected BOM
Officially supports up to four Max-Q 300 W GPUs. CPU, RAM and SSD are proposed. DeepSeek: about 188.83 GiB of Engram in RAM; engine and speed require testing. GPU support does not mean the model has been tested.
GPU support is published by the manufacturer. The selected model has not yet been tested.
4 × 2600 W Titanium; usable power depends on redundancy
CBK8 requires EPYC 9005 and Dual Rotor High Performance Fan C5WW. The specific CPU/RAM/SSD choices are our configuration enquiry, not a fixed factory SKU.
GPU support is published by the manufacturer. The selected model has not yet been tested.
Official CF3F BOM: 4 × BlueField-3 B3140H 400G and 1 × B3220 2 × 200G, CBK8 GPUs. Main components are fixed in CTO. 768 GB is the sum of GPU memory, not a single memory bank. GLM files are about 703.74 GiB: runtime headroom and engine must be checked separately.
GPU support is published by the manufacturer. The selected model has not yet been tested.
List the questions we should clarify before signing the contract.
Your AI
I suggest checking three points:
1. What is included in the price and what is charged separately.
2. The delivery deadline and handover process.
3. How both parties agree changes.
Example AI interactionAsk your question
START WITH YOUR TASK
Just two simple questions.
Tell us what AI should help with and who will use it. We’ll select the rest.
The agreed computer or server, an installed AI model and the required software. We help you get started. The quotation specifies the exact configuration, setup, price and support.
Do I need to understand models and graphics cards?
Just tell us what you want AI to help with. We’ll clarify your tasks, select a model and explain suitable options in plain language.
What is included in the price?
The catalogue shows a preliminary price for the complete system: hardware and allowances for assembly, initial setup and testing. Handover is in China. Shipping, import and complex integrations are calculated separately. Price, lead time and warranty are confirmed in the quotation before ordering.
How quickly will AI respond?
We test the selected system on agreed tasks, measuring time to first response and generation speed. For teams, we test simultaneous requests. Measurements from another system are not presented as the speed of your purchase.
Will my documents stay with me?
A local model can process documents on your own hardware. We agree deployment, staff access and connected services to meet your data requirements.
Can I choose DeepSeek or another model?
Yes. We consider DeepSeek, Qwen, GLM, MiniMax and other models. The exact version and tasks affect the hardware configuration. If you already know what you need, open the technical selector below.
Can I collect the system in China myself?
Self-collection can be arranged. The handover location, configuration, checks, documents and warranty are set out in the quotation. Shipping can also be discussed for your order.
I know the model and specificationsTechnical selection, hardware configurations and benchmark sources
FOR TECHNICAL USERS
Model, workload, hardware.
Preliminary sizing for the exact model version. Final configuration and speed are confirmed by testing.
Generations at the same time, not the number of staff or agents.
Weight format and context-memory format are separate settings.
Preliminary calculation for the exact checkpoint. Speed, actual memory use and the selected workload have not yet been measured. The native context limit is not confirmed here; the selected length needs testing.
One model replica. Context and the selected workload are subject to testing.
Group: 336 GBGPU weights: 305.79 GiB
Configuration and terms
CPU
Threadripper PRO, 24–32 cores, e.g. 9965WX/9975WX; WRX90 with direct PCIe lanes. A factory AMD EPYC server with verified lane layout is also possible.
RAM
256 GB ECC, 8 × 32 GB, 8 channels
Storage
TLC NVMe: 2 TB for system + 4 TB for models
Power
At least 3000 W usable PSU capacity; server airflow or factory liquid cooling for four cards
DeepSeek requires a specified Engram-to-RAM offload method; peak memory must be tested. GLM Flash leaves limited headroom after loading weights. Workload and placement on each GPU require testing.
Threadripper PRO, 24–32 cores, e.g. 9965WX/9975WX; WRX90 with direct PCIe lanes. A factory AMD EPYC server with verified lane layout is also possible.
RAM
256 GB ECC, 8 × 32 GB, 8 channels
Storage
TLC NVMe: 2 TB for system + 4 TB for models
Power
At least 3000 W usable PSU capacity; server airflow or factory liquid cooling for four cards
DeepSeek requires a specified Engram-to-RAM offload method; peak memory must be tested. GLM Flash leaves limited headroom after loading weights. Workload and placement on each GPU require testing.
Studio 288
4 × RTX PRO 5000 Blackwell 72 GB
Full configuration
Processor
Threadripper PRO, 24–32 cores, WRX90, four CPU-connected PCIe slots
RAM
256 GB ECC, 8 × 32 GB
Storage
NVMe TLC: 2 TB + 4 TB
Power
Factory 2050–2200 W power; four 300 W cards with verified clearance for active cooling
GLM Flash is excluded. Four Qwen replicas require four single-GPU groups; speed must not be multiplied automatically.
Flow 192
2 × RTX PRO 6000 Blackwell Workstation 96 GB
Full configuration
Processor
Threadripper PRO, 24 cores / WRX90, two direct PCIe x16 slots
RAM
256 GB ECC, 8 × 32 GB
Storage
NVMe TLC: 2 TB + 4 TB
Power
Factory 2050–2200 W power for two 600 W cards
Flash-Next leaves no more than 19.22 GiB nominally before buffers.
Core 72
1 × RTX PRO 5000 Blackwell 72 GB
Full configuration
Processor
Ryzen 9 9950X, motherboard with a full CPU-connected PCIe x16 slot
RAM
64–128 GB DDR5, motherboard-compatible UDIMM kit
Storage
TLC NVMe: 1–2 TB for system + 4 TB for models
Power
Quality 1000 W PSU, ventilated case, actively cooled GPU
Qwen BF16 passes the preliminary weight-size check. BF16 speed and C4/C8 modes have not been measured.
Core Duo
2 × RTX PRO 5000 Blackwell 72 GB
Full configuration
Processor
Threadripper PRO, 24 cores, WRX90, two direct PCIe x16 slots
RAM
128 GB ECC, 4 × 32 GB; expandable to 256 GB
Storage
NVMe TLC: 2 TB + 4 TB
Power
Quality 1600 W PSU, two actively cooled GPUs
144 GB is not used as a single pool for Flash-Next.
Core Twin
2 workstations, each with 1 × RTX PRO 5000 72 GB
Full configuration
Processor
Ryzen 9 9950X in each, separate OS / containers and code-build resources
RAM
64–128 GB DDR5 UDIMM in each
Storage
1–2 TB for system + 4 TB for models in each
Power
1000 W PSU and verified active cooling in each
Adding the VRAM of two workstations does not permit splitting one model across them without a separate deployment recipe.
Radeon 64
2 × Radeon AI PRO R9700 32 GB
Full configuration
Processor
Ryzen 9 9950X; motherboard with two CPU-connected slots and confirmed x8/x8 mode
RAM
128 GB DDR5 UDIMM
Storage
NVMe TLC: 1–2 TB + 4 TB
Power
1200–1600 W PSU, cooling for two 300 W cards
Weight placement, ROCm, TP2, P2P and concurrent requests are untested.
Radeon 128
4 × Radeon AI PRO R9700 32 GB
Full configuration
Processor
Threadripper PRO, 24 cores, WRX90, four direct PCIe slots, ROCm-approved platform
RAM
128 GB ECC, 4 × 32 GB
Storage
NVMe TLC: 2 TB + 4 TB
Power
Factory 2050–2200 W power, four 300 W GPUs
Simultaneous operation of both TP2 groups has not been tested.
External benchmarks and test conditions
These are the authors’ results on different hardware. They cannot be applied to catalogue systems.
DeepSeek V4.1 Flash
4 × RTX PRO 6000 Workstation 96 GB, Threadripper PRO 9985WX, approximately 639 GiB RAM, Engram in RAM, patched SGLang and DSpark.
Price on requestConfiguration subject to manufacturer confirmation · 2026-10-13.1
Model to test: GLM-5.3-Flash
Selected configuration
2 × AMD EPYC 9535 · 64 cores each 768 GB ECC RDIMM · 24 × 32 GB 2 × 960 GB M.2 + 2 × 3,84 TB NVMe 4 × 2600 W Titanium; usable power depends on redundancy
CBK8 requires EPYC 9005 and Dual Rotor High Performance Fan C5WW. The specific CPU/RAM/SSD choices are our configuration enquiry, not a fixed factory SKU.
Availability, price, supply in China and handover terms are not yet confirmed.
