MiMo-V2.5-Pro hardware requirements
Needs at least 4 GPUs at FP8 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 32 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $16.00/hr.
Top options for MiMo-V2.5-Pro
The strongest part each vendor offers for MiMo-V2.5-Pro at FP8 with 32,768 tokens per user, rentable parts first. Each card is one replica, the fewest accelerators that hold the model, and the line it draws on the graph below. Select a card to plot it.
Tokens per second as users grow
Each line is one replica, the fewest accelerators that hold the model, serving more users at once. It climbs while memory bandwidth is the limit, flattens once compute is, and stops where memory runs out at 32,768 tokens per user. Hover a line to bring it forward.
8 x B300 SXM 262GB: peaks at 4.4k tokens per second when memory fills at 580 users.8 x Instinct MI325X OAM: peaks at 2.7k tokens per second when memory fills at 551 users.16 x Ascend 910C: peaks at 1.7k tokens per second when memory fills at 546 users.16 x Gaudi 2 96GB: peaks at 1.5k tokens per second when memory fills at 258 users.8 x TPU v7 192GB: peaks at 3.5k tokens per second when memory fills at 263 users.
Memory fills at: 8 x B300 SXM 262GB 580 users, 8 x Instinct MI325X OAM 551 users, 16 x Ascend 910C 546 users, 16 x Gaudi 2 96GB 258 users, 8 x TPU v7 192GB 263 users. Roofline estimates at FP8.
Bare minimum
- Tokens/s
- 504
- Per hour
- $23.96
- Per million tokens
- $13.20
- Memory used
- 92%
- Tokens/s
- 375
- Per hour
- $8.00
- Per million tokens
- $5.92
- Memory used
- 70%
- Tokens/s
- 659
- Per hour
- $8.00
- Per million tokens
- $3.37
- Memory used
- 73%
Configurations
Every datacenter GPU in the catalog, sized for these settings. Change precision, context and concurrency and the whole page follows. Each point is one replica: the smallest tensor-parallel group of that GPU that holds the model.
4 x Instinct MI355X OAM: 940 tokens/s per replica at $23.96/hr.4 x GB300 NVL72 GPU 288GB: 938 tokens/s per replica at $49.68/hr.8 x GB200 NVL72 GPU 186GB: 1.8k tokens/s per replica at $84.00/hr.4 x Instinct MI350X OAM: 921 tokens/s per replica at $24.64/hr.8 x B300 SXM 262GB: 1.7k tokens/s per replica at $39.92/hr.8 x B200 SXM 180GB: 1.7k tokens/s per replica at $32.72/hr.8 x Instinct MI325X OAM: 1.2k tokens/s per replica at $30.40/hr.8 x Instinct MI300X 192GB: 1.1k tokens/s per replica at $19.12/hr.8 x H200 SXM 141GB: 951 tokens/s per replica at $19.84/hr.16 x H100 SXM5 80GB: 1.3k tokens/s per replica at $29.28/hr.8 x H200 NVL 141GB: 899 tokens/s per replica at $28.91/hr.16 x H100 NVL 94GB: 1.4k tokens/s per replica at $51.04/hr.16 x H100 PCIe 80GB: 814 tokens/s per replica at $40.00/hr.16 x RTX PRO 6000 Blackwell Workstation Edition: 683 tokens/s per replica at $28.80/hr.16 x RTX PRO 6000 Blackwell Server Edition: 625 tokens/s per replica at $11.04/hr.16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition: 659 tokens/s per replica at $8.00/hr.16 x Gaudi 2 96GB: 814 tokens/s per replica at $17.60/hr.16 x A100 PCIe 80GB: 455 tokens/s per replica at $21.60/hr.16 x A100 SXM4 80GB: 466 tokens/s per replica at $22.40/hr.
Priced configurations only; 28 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.
| Rent | |||
|---|---|---|---|
| 4 | 2.9k tok/s | Specs → | |
| 4 | 2.7k tok/s | Specs → | |
TPU 8t announced | 8 | ~1.4k tok/s | Specs → |
TPU 8i announced | 4 | ~801 tok/s | Specs → |
| 4 | 940 tok/s | Rent → | |
| 4 | 938 tok/s | Rent → | |
| 8 | 1.8k tok/s | Rent → | |
| 4 | 921 tok/s | Rent → | |
| 8 | 1.6k tok/s | Specs → | |
Instinct MI430X OAM announced | 4 | ~1.2k tok/s | Specs → |
| 8 | 1.7k tok/s | Rent → | |
| 8 | 1.7k tok/s | Rent → | |
| 4 | 1.3k tok/s | Specs → | |
| 8 | 1.6k tok/s | Specs → | |
| 8 | ~2.4k tok/s | Specs → | |
| 8 | 1.2k tok/s | Rent → | |
| 8 | 1.1k tok/s | Rent → | |
| 4 | 771 tok/s | Specs → | |
| 8 | 951 tok/s | Rent → | |
H100 SXM5 80GB 2 nodes | 16 | 1.3k tok/s | Rent → |
Data Center GPU Max 1550 128GB 2 nodes | 16 | ~962 tok/s | Specs → |
| 8 | ~899 tok/s | Rent → | |
H100 NVL 94GB 2 nodes | 16 | ~1.4k tok/s | Rent → |
| 16 | ~930 tok/s | Specs → | |
H100 PCIe 80GB 2 nodes | 16 | 814 tok/s | Rent → |
H800 PCIe 80GB 2 nodes | 16 | 814 tok/s | Specs → |
| 8 | 668 tok/s | Specs → | |
| 16 | 683 tok/s | Rent → | |
| 16 | 626 tok/s | Specs → | |
| 16 | ~626 tok/s | Specs → | |
| 16 | 625 tok/s | Rent → | |
Gaudi 3 128GB 2 nodes | 16 | 1.3k tok/s | Specs → |
TPU v5p 95GB 2 nodes | 16 | 661 tok/s | Specs → |
| 16 | 659 tok/s | Rent → | |
| 16 | ~225 tok/s | Specs → | |
Gaudi 2 96GB 2 nodes | 16 | 814 tok/s | Rent → |
| 16 | 536 tok/s | Specs → | |
MI250X 128GB 2 nodes | 16 | ~641 tok/s | Specs → |
Instinct MI250 128GB 2 nodes | 16 | ~619 tok/s | Specs → |
A800 PCIe 80GB 2 nodes | 16 | ~455 tok/s | Specs → |
A800 SXM4 2 nodes | 16 | ~466 tok/s | Specs → |
A100 PCIe 80GB 2 nodes | 16 | ~455 tok/s | Rent → |
A100 SXM4 80GB 2 nodes | 16 | ~466 tok/s | Rent → |
RTX PRO 5000 Blackwell 72GB 2 nodes | 16 | ~346 tok/s | Specs → |
| 8 | 327 tok/s | Specs → | |
H20 96GB 2 nodes | 16 | 577 tok/s | Specs → |
| 16 | ~134 tok/s | Specs → |
By vendor
At FP8 the NVIDIA pick is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 659 tokens/s for $8.00/hr; the AMD pick is 8 x Instinct MI300X 192GB at about 1.1k tokens/s for $19.12/hr. Per rental dollar NVIDIA delivers 1.42x the tokens of AMD here.
NVIDIA: 82.4 tokens/s per dollar, 137 tokens/s per kW (16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition).AMD: 57.9 tokens/s per dollar, 184 tokens/s per kW (8 x Instinct MI300X 192GB).Huawei: no live price, no power figure (8 x Ascend 950DT).Intel: 46.2 tokens/s per dollar, 84.8 tokens/s per kW (16 x Gaudi 2 96GB).Google: no live price, no power figure (8 x TPU v7 192GB).Other: no live price, no power figure (16 x MTT S5000).
Fleet what-if
Compare whole fleets at these settings: a thousand of one part against a hundred of another. Each fleet splits into replicas of its tensor-parallel width; GPUs left over sit idle.
| Fleet | Tokens/s |
|---|---|
1,000 x H200 SXM 141GB | 119k tok/s |
104 x GB200 NVL72 GPU 186GB | 23k tok/s |
1,000 x H200 SXM 141GB: 119k tokens/s, $2,480/hr, 700 kW.104 x GB200 NVL72 GPU 186GB: 23k tokens/s, $1,092/hr, no power figure.
H200 SXM 141GB: 119k tokens/s at 1000 GPUs.GB200 NVL72 GPU 186GB: 222k tokens/s at 1000 GPUs.
Size MiMo-V2.5-Pro for a tokens per second target
The calculator starts from your traffic instead of a fleet: set a target rate and read the replica count for every GPU.
Frequently asked questions
How much GPU memory does MiMo-V2.5-Pro need?
The weights take 1,023 GB at the native FP8 precision (2,046 GB at BF16, 1,023 GB at FP8, 576 GB at INT4). Each concurrent stream adds 350 KB of KV cache per token: 0.2 GB at 4,096 tokens and 1.7 GB at 32,768 tokens.
What is the bare minimum to run MiMo-V2.5-Pro?
4 x Instinct MI355X OAM at FP8 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00 per hour.
How many H100s do you need to run MiMo-V2.5-Pro?
16 H100s at FP8 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 16 H100s, and 1,000 tokens per second takes 16 H100s across 1 replicas.
How much does MiMo-V2.5-Pro cost per million tokens on an H100?
About $6.29 per million output tokens at FP8, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.83 per GPU-hour. The cheapest part per token is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $3.37.
What precision does MiMo-V2.5-Pro ship in?
The published checkpoint is FP8. FP8 halves the BF16 footprint with the accuracy the lab validated; INT4 is a further community quantisation.
All figures are roofline estimates from the model's config.json and each GPU's published memory, bandwidth and tensor-core peak, with fixed efficiency factors. Prices are the cheapest live on-demand listing per GPU when the site has one. They are a planning floor; a tuned serving stack can do better. 60 of 70 layers attend over a 128-token sliding window, so the cache grows only in the 10 full layers. Keys are 192 wide and values 128, counted exactly. Ships in FP8. Active parameters per the Xiaomi model card.