Eight H100 80GB GPUs can contain Kimi K2.5’s official checkpoint by aggregate file size—but the remaining margin is small enough that “fits” should not be confused with “deploys comfortably.”
The official repository lists 595.178GB of safetensors. Eight 80GB GPUs total 640GB, leaving 44.822GB before CUDA contexts, framework state, communication buffers, KV cache, multimodal processing, and placement imbalance. If the remainder divided perfectly, that would average only 5.603GB per GPU.
The boundary in four numbers
| Quantity | Value |
|---|---|
| Official checkpoint files | 595.178GB |
| 8×H100 80GB aggregate | 640GB |
| Aggregate remainder | 44.822GB |
| Average remainder under perfect balance | 5.603GB per GPU |
Why the arithmetic is not enough
Kimi K2.5 is a multimodal mixture-of-experts model. Routed experts may partition differently from attention, shared experts, the vision tower, projector, and language-model head. A deployment can therefore run out of memory on one H100 even when the sum across eight GPUs exceeds the files.
The engine must also support Kimi K2.5’s mixed compressed-tensors format. Long context, batch size, concurrency, image/video processing, temporary tensors, and communication buffers all consume the remaining memory.
Before calling 8×H100 viable
- Pin the exact model revision and serving-engine build.
- Record PCIe or SXM identity, topology, MIG state, and memory exposed per device.
- Record parallelism, component placement, and any CPU or storage offload.
- Load with short text context, batch 1, and concurrency 1.
- Measure peak memory on every device while loading and generating.
- Raise context and concurrency gradually; test multimodal input separately.
- Report correctness, failures, latency, throughput, throttling, and wall power.
Full answer: read the eight-H100 Kimi K2.5 capacity and validation guide.
Continue planning
- Check the full checkpoint arithmetic
- Compare 640GB–864GB hardware paths
- Verify H100 variants and platform requirements
- Browse cited benchmark evidence
- Stage a reproducible result
- Follow new evidence by RSS
Primary sources
- Official Kimi K2.5 model card
- Official repository API with file sizes
- Official Kimi K2.5 configuration
- NVIDIA H100 specifications
This is a capacity analysis, not a measured Kimi K2.5 performance result. Sources checked October 25, 2026.