Kimi K3 hardware requirements
Needs at least 8 GPUs at INT4 / FP4 (GB300 NVL72 GPU 288GB). For 1,000 tokens/s at 32k context you need 8 x B300 SXM 262GB at $39.92/hr.
Top options for Kimi K3
The strongest part each vendor offers for Kimi K3 at INT4 / FP4 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 GB300 NVL72 GPU 288GB: peaks at 4.6k tokens per second when memory fills at 488 users.8 x Instinct MI355X OAM: peaks at 3.7k tokens per second when memory fills at 488 users.16 x Ascend 950DT: peaks at 1.7k tokens per second when memory fills at 481 users.16 x Gaudi 3 128GB: peaks at 1.4k tokens per second when memory fills at 297 users.16 x TPU v7 192GB: peaks at 2.3k tokens per second and still has memory at 1024 users.
Memory fills at: 8 x GB300 NVL72 GPU 288GB 488 users, 8 x Instinct MI355X OAM 488 users, 16 x Ascend 950DT 481 users, 16 x Gaudi 3 128GB 297 users, 16 x TPU v7 192GB beyond 1,024 users. Roofline estimates at INT4 / FP4.
Bare minimum
- Tokens/s
- 685
- Per hour
- $110
- Per million tokens
- $44.49
- Memory used
- 71%
- Tokens/s
- 484
- Per hour
- $30.40
- Per million tokens
- $17.45
- Memory used
- 79%
- Tokens/s
- 1.4k
- Per hour
- $39.92
- Per million tokens
- $7.65
- Memory used
- 80%
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.
8 x Instinct MI355X OAM: 1.4k tokens/s per replica at $47.92/hr.8 x GB300 NVL72 GPU 288GB: 1.5k tokens/s per replica at $110/hr.16 x GB200 NVL72 GPU 186GB: 2.4k tokens/s per replica at $168/hr.8 x Instinct MI350X OAM: 1.3k tokens/s per replica at $49.28/hr.8 x B300 SXM 262GB: 1.4k tokens/s per replica at $39.92/hr.16 x B200 SXM 180GB: 2.3k tokens/s per replica at $65.44/hr.8 x Instinct MI325X OAM: 503 tokens/s per replica at $30.40/hr.16 x Instinct MI300X 192GB: 844 tokens/s per replica at $38.24/hr.16 x H200 SXM 141GB: 686 tokens/s per replica at $47.84/hr.16 x H200 NVL 141GB: 614 tokens/s per replica at $57.82/hr.
Priced configurations only; 20 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.
| Rent | |||
|---|---|---|---|
| 4 | 2.2k tok/s | Specs → | |
| 8 | 3.9k tok/s | Specs → | |
TPU 8t announced | 8 | ~1.2k tok/s | Specs → |
TPU 8i announced | 8 | ~1.4k tok/s | Specs → |
| 8 | 1.4k tok/s | Rent → | |
| 8 | 1.5k tok/s | Rent → | |
GB200 NVL72 GPU 186GB 2 nodes | 16 | 2.4k tok/s | Rent → |
| 8 | 1.3k tok/s | Rent → | |
TPU v7 192GB 2 nodes | 16 | ~1.3k tok/s | Specs → |
Instinct MI430X OAM announced | 4 | ~1.4k tok/s | Specs → |
| 8 | 1.4k tok/s | Rent → | |
B200 SXM 180GB 2 nodes | 16 | 2.3k tok/s | Rent → |
| 8 | 1.9k tok/s | Specs → | |
B100 SXM 192GB 2 nodes | 16 | 2.2k tok/s | Specs → |
| 8 | ~2.4k tok/s | Specs → | |
| 8 | ~503 tok/s | Rent → | |
Instinct MI300X 192GB 2 nodes | 16 | ~844 tok/s | Rent → |
| 8 | 1.1k tok/s | Specs → | |
H200 SXM 141GB 2 nodes | 16 | ~686 tok/s | Rent → |
Data Center GPU Max 1550 128GB 2 nodes | 16 | ~533 tok/s | Specs → |
H200 NVL 141GB 2 nodes | 16 | ~614 tok/s | Rent → |
| 16 | ~514 tok/s | Specs → | |
| 16 | 909 tok/s | Specs → | |
| 16 | 482 tok/s | Specs → | |
Gaudi 3 128GB 2 nodes | 16 | ~805 tok/s | Specs → |
| 16 | ~147 tok/s | Specs → | |
| 16 | 411 tok/s | Specs → | |
MI250X 128GB 2 nodes | 16 | ~316 tok/s | Specs → |
Instinct MI250 128GB 2 nodes | 16 | ~303 tok/s | Specs → |
H20 141GB HBM3e 2 nodes | 16 | ~147 tok/s | Specs → |
By vendor
At INT4 / FP4 the NVIDIA pick is 8 x B300 SXM 262GB at about 1.4k tokens/s for $39.92/hr; the AMD pick is 8 x Instinct MI355X OAM at about 1.4k tokens/s for $47.92/hr. Per rental dollar NVIDIA delivers 1.27x the tokens of AMD here.
NVIDIA: 36.3 tokens/s per dollar, 129 tokens/s per kW (8 x B300 SXM 262GB).AMD: 28.6 tokens/s per dollar, 122 tokens/s per kW (8 x Instinct MI355X OAM).Huawei: no live price, no power figure (16 x Ascend 950DT).Intel: no live price, 55.9 tokens/s per kW (16 x Gaudi 3 128GB).Google: no live price, no power figure (16 x TPU v7 192GB).Other: no live price, 61.1 tokens/s per kW (16 x Cloud AI 100 Ultra).
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 |
|---|---|
1,008 x H200 SXM 141GB | 43k tok/s |
112 x GB200 NVL72 GPU 186GB | 17k tok/s |
1,008 x H200 SXM 141GB: 43k tokens/s, $3,014/hr, 706 kW.112 x GB200 NVL72 GPU 186GB: 17k tokens/s, $1,176/hr, no power figure.
H200 SXM 141GB: 43k tokens/s at 1008 GPUs.GB200 NVL72 GPU 186GB: 152k tokens/s at 1008 GPUs.
Size Kimi K3 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 Kimi K3 need?
The weights take 1,564 GB at the native INT4 / FP4 precision (5,560 GB at BF16, 2,780 GB at FP8, 1,564 GB at INT4). Each concurrent stream adds 27 KB of KV cache per token: 0.1 GB at 4,096 tokens and 0.9 GB at 32,768 tokens.
What is the bare minimum to run Kimi K3?
8 x GB300 NVL72 GPU 288GB at INT4 / FP4 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 8 x Instinct MI325X OAM at $30.40 per hour.
How many H100s do you need to run Kimi K3?
more than sixteen H100s at INT4 / FP4 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is more than sixteen H100s.
How much does Kimi K3 cost per million tokens on an H100?
No live H100 rental price is on file right now, so the cost per token cannot be quoted. The configuration table lists every part that has one.
What precision does Kimi K3 ship in?
The published checkpoint is MXFP4. That 4-bit format is what the lab validated, so the INT4 column is the native one; BF16 figures describe a dequantised copy nobody would deploy.
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. Kimi Linear hybrid: 69 KDA linear-attention layers hold a fixed 434 MB of FP32 state per sequence, and only 24 layers keep an MLA cache (27 KB per token). Ships in MXFP4; 896 routed experts at a reduced width, 16 per token. Active parameters per the Moonshot model card.