Short answer: the official DeepSeek-V3.2 and DeepSeek-V3.2-Speciale checkpoints each total 689.483GB. Eight 80GB accelerators are 49.483GB short before runtime memory. The first defensible hardware conversation begins with qualified platforms providing substantially more than 689.483GB in aggregate—such as 8×96GB, 6×141GB, 5×180GB, 3×270GB documented usable, 4×192GB, 3×256GB or 3×288GB—then verifies per-device placement, topology, software and workload headroom.
Do not buy from aggregate arithmetic alone
A cluster total above 689.483GB proves only that the files clear nominal aggregate capacity. It does not prove that shards balance, the runtime supports the topology, live allocations fit on every device or the model performs acceptably. Compare complete qualified systems and require a measured workload.
Start with the DeepSeek-V3.2 capacity guide for the exact repository totals and the multi-GPU guide for placement and pooling constraints.
DeepSeek-V3.2 hardware capacity table
| Candidate platform path | Nominal aggregate memory | File-only difference from 689.483GB | Capacity interpretation |
|---|---|---|---|
| 8× H100/A100 80GB | 640GB | 49.483GB short | Official files do not fit wholly |
| 9× 80GB accelerators | 720GB | 30.517GB margin | Arithmetic clears files; too tight to recommend without a proven platform-specific plan |
| 4× B200 at 180GB | 720GB | 30.517GB margin | Same tight aggregate boundary; Blackwell software and per-device placement remain gates |
| 5× H200 141GB | 705GB | 15.517GB margin | Unsafe file-size boundary, not a responsible deployment target |
| 8× H100 NVL 94GB | 752GB | 62.517GB margin | Files clear aggregate capacity; exact NVL system topology and headroom still required |
| 8× RTX PRO 6000 Blackwell 96GB | 768GB | 78.517GB margin | Workstation-card path; physical layout, host, peer communication and software must be proven |
| 4× AMD MI300X 192GB | 768GB | 78.517GB margin | Qualified OAM server path; ROCm and Infinity Fabric placement must be proven |
| 3× AMD MI325X 256GB | 768GB | 78.517GB margin | Fewer large-memory devices; still requires a supported three-device placement plan |
| 3× B300 at 270GB documented usable | 810GB | 120.517GB margin | Blackwell Ultra capacity path; exact platform exposure and current software support are gates |
| 6× H200 141GB | 846GB | 156.517GB margin | More file headroom than 8×96GB; complete HGX/DGX topology and workload proof required |
| 3× AMD MI350X/MI355X 288GB | 864GB | 174.517GB margin | CDNA 4 capacity path; MI350X air-cooled and MI355X liquid-cooled systems are not interchangeable |
| 5× B200 180GB | 900GB | 210.517GB margin | More file headroom; platform cost, power and measured need remain separate decisions |
Best starting point by deployment constraint
- Existing 8-GPU 96GB platform: 768GB clears the files by 78.517GB. This is a plausible lower-bound test platform only when the runtime, topology and per-device peaks are already validated.
- Hopper continuity: six H200 141GB devices provide 846GB and 156.517GB of file-only margin. Five H200s leave only 15.517GB and should not be presented as a comfortable target.
- Blackwell data center: five B200 devices at 180GB provide 900GB. Four B200s provide only a 30.517GB file margin. Three B300 devices at the documented 270GB usable figure provide 810GB, subject to exact platform exposure and software support.
- AMD Instinct: four MI300X, three MI325X or three MI350X/MI355X cross the file boundary. The decision then moves to qualified server availability, ROCm target support, fabric topology, cooling and measured workload behavior.
- Workstation cards: eight RTX PRO 6000 Blackwell 96GB cards reach 768GB in arithmetic, but this is not automatically equivalent to an eight-accelerator HGX or Instinct platform. Slot spacing, power, host lanes, peer communication and runtime support can rule it out.
Why the smallest fitting count is often the wrong target
A 720GB or 705GB total may clear the repository files while leaving only 30.517GB or 15.517GB across the entire system. Runtime allocations are per device, not one shared reserve. Cache, activations, communication buffers, graph capture, allocator behavior and uneven expert or tensor placement can exhaust one device even while aggregate arithmetic looks positive. Long context and concurrency make the gap larger.
CPU or storage offload can reduce accelerator capacity requirements, but it changes placement and performance. State the offloaded tensors, host RAM, transfer path and observed behavior; do not relabel an offloaded deployment as wholly accelerator resident.
NVIDIA platform checks
- Confirm exact PCIe, SXM, NVL, HGX or DGX system identity and memory exposed per accelerator.
- Record MIG state, NVLink/NVSwitch topology, driver, CUDA, container, framework and DeepSeek-specific kernels.
- For Blackwell and Blackwell Ultra, verify current compute-target and communication-library support rather than assuming a Hopper image carries forward.
- Compare complete server power, cooling, host CPUs, system RAM, storage, networking, service and delivery—not loose accelerator modules.
AMD Instinct platform checks
- Confirm exact MI300X, MI325X, MI350X or MI355X server SKU, exposed memory and Infinity Fabric topology.
- Pin the ROCm release, gfx target, container, framework, quantization/attention kernels and communication libraries.
- Distinguish MI350X’s air-cooled platform path from MI355X’s higher-power direct-liquid-cooled path.
- Measure correctness as well as speed; a load or token count is not enough if kernels fall back or output differs.
Reproducible buying test
- Pin DeepSeek-V3.2 or V3.2-Speciale and the exact repository revision.
- Record every device, exposed memory, topology, host and placement strategy.
- Pin runtime, build, driver, kernels, launch flags and any offload.
- Fix input/output lengths, context setting, cache precision, batch and concurrency.
- Measure per-device peak memory, host-memory peak, time to first token, prompt processing, generation and wall power.
- Repeat the target workload and inspect output correctness and failures before comparing price or operating cost.
Compare NVIDIA data-center tiers, compare AMD Instinct tiers, check source-specific measurements, build a reproducible record, or save private hardware criteria.
Primary sources
- Official DeepSeek-V3.2 repository API with blob sizes
- Official DeepSeek-V3.2-Speciale repository API with blob sizes
- NVIDIA HGX H100/H200/B200 component reference
- NVIDIA current HGX component reference
- NVIDIA Blackwell Ultra platform datasheet
- NVIDIA RTX PRO 6000 Blackwell Workstation Edition datasheet
- AMD Instinct MI300X specifications
- AMD Instinct MI325X specifications
- AMD ROCm MI350 architecture reference
- AMD MI355X system-acceptance reference
Reviewed October 25, 2026. Capacity arithmetic and manufacturer specifications are not independent DeepSeek performance benchmarks. HighVRAM has not reproduced these platform configurations.