MiniMax-M2.5’s official repository currently lists 125 mapped safetensors files totaling 230.134GB. That means a nominal 240GB accelerator pool leaves only 9.866GB before the runtime allocates KV cache, workspaces, communication buffers or other live state. The files fit by subtraction; a useful resident deployment is not proven.
The capacity boundary
- 192GB: 38.134GB short.
- 240GB: 9.866GB file-only margin—unsafe as a blanket resident-fit promise.
- 256GB: 25.866GB file-only margin.
- 288GB: 57.866GB file-only margin.
The full MiniMax-M2.5 VRAM guide records the exact repository state, the 196,608-position configuration ceiling and every file-only boundary from 192GB through 384GB.
Why 240GB is a trap
The checkpoint files consume 230,134,264,560 bytes. A model server still needs non-weight allocations: KV cache, attention state, temporary workspaces, graph capture, communication buffers and allocator headroom. In a multi-device system the 9.866GB remainder is not one interchangeable pool; one shard or duplicated allocation can exhaust an individual accelerator first.
A shorter context, lower concurrency or host-memory offload may allow a constrained deployment. State those choices. They do not turn 240GB into a generally comfortable resident configuration.
The official repository has a metadata wrinkle
The current safetensors index maps 125 files numbered 00000 through 00124, but their names contain the suffix of-00126. The same index reports 466,083,044,352 bytes as logical tensor total_size, roughly twice the summed stored FP8 blobs. This article uses the observed 230.134GB download total for file-fit arithmetic and discloses the mismatch rather than silently choosing one number.
Why 10B active is not the storage requirement
The publisher describes MiniMax-M2.5 as a 230B-total, 10B-active mixture-of-experts model. Routing fewer experts for each token can affect compute, but the deployment still needs access to stored expert weights. A 10B-active label does not make the official checkpoint a 10B-model download.
256GB and 288GB are different decisions
At 256GB, the official stored files leave 25.866GB. At 288GB, they leave 57.866GB. MI325X, MI350X and MI355X are server accelerators in qualified OAM platforms, not desktop upgrades. Verify ROCm target and kernel support, Infinity Fabric topology, host memory, cooling and the exact system. NVIDIA multi-device paths require the same discipline for runtime version, tensor/expert placement and communication topology.
What a useful result must disclose
- Exact repository revision and downloaded files.
- Every accelerator, runtime-visible memory and host system.
- Interconnect, placement and any duplicated allocations.
- Runtime/build, driver, kernels and launch flags.
- Input/output tokens, context setting, cache precision, batch and concurrency.
- Any CPU, host-memory or storage offload.
- Per-device peak memory, host-memory peak, correctness, prompt processing, decode and time to first token.
Plan multi-GPU placement, check source-specific measurements, build a reproducible MiniMax record, save private hardware criteria, or follow the RSS feed.
Primary sources
- Official MiniMax-M2.5 model card
- Official repository API with blob sizes
- Official configuration
- Official safetensors index
- Official vLLM deployment guide
Published October 25, 2026. File-size arithmetic is not an independent performance benchmark.