Approximately 865 GB assistant artifact
DeepSeek-V4-Pro is the practical larger assistant candidate. Model quality still depends on the task, engine, prompt, tools, retrieval, and controls.
Infrastructure
Size the platform around the whole runtime.
Model files are only part of the memory plan. The operating system, inference engine, context, retrieval, concurrency, model staging, failover, and future growth all consume capacity.
MemoryInstalled and usable capacity documented separately
StorageEnterprise NVMe sized for model, staging, and recovery
ComputeCPU, GPU, network, power, and thermals validated together
Memory tiers
Three planning classes cover very different operating envelopes.
These are architecture classes, not fixed inventory or automatic quotes. Exact platform limits, DIMM population, matched memory, firmware, accelerator support, power, thermals, and supplier availability must be verified for the selected chassis.
| Memory class | Practical role | Large-model planning note | Required proof |
|---|---|---|---|
| 1.125 TiB installed | Focused evaluation, retrieval, document, and assistant workloads. | The approximately 865 GB mixed-precision DeepSeek-V4-Pro assistant artifact may fit with limited raw headroom. Runtime overhead, context, concurrency, and retrieval must be measured. | Exact usable memory, load test, task quality, latency, and stability. |
| 2.0–2.25 TiB installed | Growth, larger context, concurrent work, and broader model evaluation. | The 1.61 TB FP8 DeepSeek-V4-Pro-Base checkpoint may be loadable, but this is a research floor rather than the preferred operating class. | NUMA behavior, engine overhead, accelerator path, and workload benchmark. |
| 5–6 TiB installed | Frontier-scale private research, staging, adaptation, and multi-model work. | Recommended class for the optional 1.61 TB base checkpoint with room for the operating environment, staging, retrieval, and useful concurrency. | Validated homogeneous memory population, burn-in, recovery test, and task acceptance. |
Large-model option
The 1.61 TB option is a foundation checkpoint—not a finished assistant.
DeepSeek-V4-Pro-Base is an open-weight base model for controlled research or adaptation. The ready assistant release uses a smaller mixed-precision artifact of roughly 865 GB. Neither should be represented as equivalent to a proprietary hosted product without same-task evidence.
1.61 TB FP8 base artifact
DeepSeek-V4-Pro-Base provides a frontier-scale foundation checkpoint for evaluation and adaptation, not a ready business assistant.
Same sanitized tasks
Compare both candidates and the current baseline on completion quality, source fidelity, correction rate, latency, throughput, and operating cost.
Model identity and artifact sizes should be rechecked against the official DeepSeek-V4-Pro and DeepSeek-V4-Pro-Base repositories before a final build.
Whole-platform review
A server is a system, not a memory total.
A final architecture ties capacity to reliability, serviceability, and the response target.
Chassis and memory population
Exact CPU, DIMM type, rank, population rules, firmware, NUMA layout, and usable memory are documented before procurement.
Boot, models, staging, and recovery
Mirrored boot and enterprise NVMe are separated by role, with capacity and endurance sized for model artifacts, swap policy, indexes, backups, and updates.
GPU and interconnect
GPU memory, PCIe lanes, peer links, power, cooling, and engine support are sized to the accepted latency and throughput target.
Network, power, and service
Redundant power, rack fit, remote management, network path, monitoring, backups, spare strategy, and recovery ownership are part of the design.
Start with the actual job
Get the workload and the hardware in the same room.
A planning call should establish model candidates, data boundary, required response time, concurrency, available power and rack conditions before a supplier quote is treated as final.