On this page
To run Kimi K3.1 Pro you need at least 1.56 terabytes of GPU memory. That is the size of the weights Moonshot AI put on Hugging Face by 5 November 2026. One AWS machine with eight H100s holds 640 GB, so two of them come up 281 GB short. The arithmetic floor is three H100 machines, or two with H200s.
If you want Kimi K3.1 Pro on rented GPUs, plan for three 8x H100 machines or two 8x H200 machines, and expect the serving engines to ask for more.
If one 8x H100 machine is your budget, DeepSeek-V4.1-Flash and MiniMax M3.1 Flash both fit on it.
If you want a desk box, none of the three fits on a DGX Spark, and only DeepSeek has a published setup for a DGX Station.
The short answer
GPU memory is the memory on the graphics cards themselves. The whole model has to sit in it, plus working room for every conversation in progress. Here is the floor for each model, set by file size alone. Below, Kimi, DeepSeek and MiniMax mean these three versions.
| Kimi K3.1 Pro | DeepSeek-V4.1-Flash | MiniMax M3.1 Flash | |
|---|---|---|---|
| Size | 2.8T parameters, 104B used per token | 552B parameters, 8B to 16B used per token | About 428B parameters, 23B used per token |
| Weights on Hugging Face, 11 Nov 2026 | 1,561 GB | 510 GB | 252 GB |
| Floor plus 20% headroom | 1,874 GB | 612 GB | 303 GB |
| Smallest 8x H100 setup that holds it | 3 machines | 1 machine | 1 machine |
| Smallest 8x H200 setup | 2 machines | 1 machine | 1 machine |
| Serving engines' own guidance | 8 Blackwell GPUs (vLLM); 32 H100s or 16 H200s as unverified layouts (SGLang) | 8 H100s or 4 H200s (vLLM) | 8 H100s or H200s, a full machine (vLLM) |
Weight sizes are the sums of each repository's files, read from the Hugging Face listing for Kimi K3.1 Pro, DeepSeek-V4.1-Flash and MiniMax M3.1 Flash on 11 November 2026. The 20% headroom is the factor vLLM's DeepSeek recipe uses for its minimum. These are floor estimates from file size, not configurations we tested.
Why does file size set the floor?
Because a model reads all of its weights from GPU memory, and they can't be fetched from disk fast enough mid-answer. All three models are mixtures of experts: each word-piece uses only a slice of the model, but any slice can be picked next, so every slice has to be loaded.
All three also ship already compressed. Most of each model is stored at 4 bits a number and the rest at 8 or 16, per Kimi's model card and the other two models' config files. So the file size is close to the memory the weights take once loaded. There is no smaller official version to fall back on.
One exception matters. About 183 GiB of DeepSeek's files are lookup tables it calls Engram memory, per vLLM's recipe. vLLM can keep those tables in the machine's ordinary memory and read them from there "with the output unchanged". That is how its recipes fit DeepSeek on four H200s.
How much working room does a conversation need?
Every conversation needs its own working memory, called the KV cache: notes the model keeps on each word-piece it has read, so it doesn't reread the whole conversation for every new word. The cache grows with the length of the conversation and with the number running at once.
A token is one word-piece, about three quarters of a word. The three models differ more here than in weights.
| Working memory | Per word-piece | One 128,000-token conversation | One 1M-token conversation | 20 agents at 128,000 tokens |
|---|---|---|---|---|
| DeepSeek-V4.1-Flash | 890 bytes | 0.11 GB | 0.89 GB | 2.3 GB |
| Kimi K3.1 Pro | about 13,800 bytes | 1.8 GB | 13.8 GB | 35 GB |
| MiniMax M3.1 Flash | about 61,400 bytes | 7.9 GB | 61 GB | 157 GB |
DeepSeek's 890 bytes is its own figure, from its model card. The other two rows are our arithmetic.
For Kimi, vLLM's recipe gives 576 bytes per word-piece for each of the 24 layers that keep a full cache. We multiplied the two, leaving out the other 69 layers, whose memory stays a fixed size however long the conversation runs. For MiniMax, we multiplied the layer and head counts in its config file by the 8-bit cache vLLM sets on H100s. Both leave out the sparse-attention index. Treat them as rough.
MiniMax has the smallest weights and the largest cache. On one H100 machine its weights leave about 388 gigabytes free, so it serves many agents, but long conversations eat that room quickly. vLLM's H100 command for it caps conversations at 262,144 tokens and 80 at once.
Which machines can hold each model?
Line the floor up against each machine's total GPU memory. A model fits when its weights plus headroom are smaller than the total.
| Machine | GPU memory | Kimi K3.1 Pro (1,561 GB) | DeepSeek (510 GB) | MiniMax (252 GB) |
|---|---|---|---|---|
| DGX Spark | 128 GB | No | No | No |
| AWS p4d.24xlarge, 8x A100 40 GB | 320 GB | No | No | Weights fit; no engine recipe |
| AWS p5.48xlarge, 8x H100 | 640 GB | 3 machines | Yes | Yes |
| Azure ND H100 v5, 8x H100 | 640 GB | 3 machines | Yes | Yes |
| Google A3 Edge, 8x H100 | 640 GB | 3 machines | Yes | Yes |
| DGX Station (GB300) | 748 GB, of which 252 GB on the GPU | No | Yes, with vLLM's offload setup | Not listed |
| AWS p5en.48xlarge, 8x H200 | 1,128 GB | 2 machines | Yes | Yes |
Memory totals come from AWS's P4 and P5 pages, Microsoft's ND H100 v5 page, Google's accelerator machine page, and NVIDIA's DGX Station and DGX Spark pages.
Kimi. Two H100 machines hold 1,280 gigabytes, 281 gigabytes less than the weights. Three hold 1,920 gigabytes, which clears the 1,874-gigabyte floor with 46 gigabytes to spare, little room for conversations.
Two H200 machines hold 2,256 gigabytes and leave about 380 gigabytes above the floor. The engines ask for more than the floor.
vLLM's recipe lists no H100 or H200 setup at all; its baseline is eight B300 GPUs in one machine, or two GB300 trays. SGLang's cookbook gives 16 H200s or 32 H100s as layouts, and says they are still in verification. Thirty-two H100s is four machines, one more than our floor.
DeepSeek. One H100 machine holds it. vLLM's recipe has a command for exactly that, with the Engram tables moved to ordinary memory. On H200 the same recipe uses four of the eight GPUs and leaves about 38.5 GiB per GPU for working memory.
MiniMax. vLLM's recipe says "H100 or H200, TP8 on a full node", meaning all eight GPUs in one machine. The weights would also fit in an A100 machine's 320 gigabytes, but neither engine lists the A100, and the A100 has no 8-bit floating-point maths, per NVIDIA's datasheet. MiniMax's format leans on that, so the A100 is untested ground.
What about a DGX Station or a DGX Spark?
A DGX Spark has 128 gigabytes, so none of the three fits; the smallest, MiniMax, is about twice that. A DGX Station quotes 748 gigabytes, but only 252 gigabytes of it is memory on the GPU, read at 7.1 TB/s. The other 496 gigabytes is the processor's memory, read at 396 gigabytes/s, per NVIDIA's Station page.
The split between the two pools decides speed. Writing each word-piece means reading the active weights from memory, so weights left in the slower pool slow every answer. vLLM's DeepSeek recipe runs on one Station by keeping the Engram tables and 60 GiB of experts in the slower pool. Kimi doesn't fit even two Stations, which add up to 1,496 gigabytes. Our DGX Spark vs DGX Station comparison covers the two boxes in full.
Can you rent these machines in Canada?
They are listed, though the H100 and H200 machines are hard to get. AWS lists p4d and p5 in Montreal, Azure lists ND H100 v5 in Canada Central, and Google lists H100 machines in Toronto. AWS also rents the H200 p5en in Montreal.
| Machine | Canadian region | On-demand price, 11 Nov 2026 | Same machine in Virginia |
|---|---|---|---|
| p4d.24xlarge, 8x A100 40 GB | AWS ca-central-1, Montreal | $25.252 an hour | $21.958 |
| p5.48xlarge, 8x H100 | AWS ca-central-1, Montreal | $63.296 an hour | $55.04 |
| p5en.48xlarge, 8x H200 | AWS ca-central-1, Montreal | $72.79 an hour | $63.30 |
| ND96isr H100 v5, 8x H100 | Azure Canada Central | $117.98 an hour | $98.32 (East US) |
| A3 Edge, 8x H100 | Google Toronto, northamerica-northeast2-c | Not read | Not read |
AWS prices are from AWS's price list, read on 11 November 2026, and match its on-demand pricing page. Azure's prices come from its public retail price list, in US dollars for Linux pay-as-you-go, read the same day. Google's comes from its GPU locations page, last updated 10 November 2026. AWS's price list shows no GPU machines in its Calgary region (ca-west-1).
What that means per model, running around the clock for a 730-hour month in Montreal:
- DeepSeek or MiniMax on one p5: $63.296 an hour, about $46,206 a month.
- Kimi on three p5s: $189.89 an hour, about $138,618 a month, at our floor, which SGLang's layout exceeds by a machine.
GPU quota starts at zero on all three clouds, so request it before you plan around a machine. Listed is not the same as available. In our experience deploying for Canadian clients, H100 and H200 capacity in Canadian regions is hard to get on demand, and the providers usually ask for a one- to three-year commitment before they allocate it. For Kimi K3.1 Pro, that means committing to three H100 machines or two H200 machines for a year or more; the monthly figures above are on-demand list prices. Large GPU machines are scarcer in Canadian regions than in American ones, per our Canadian residency guide. Our pillar on GPUs you can rent in Canada covers the smaller cards, and H100 vs A100 for LLM inference covers why the H100 machine is the one that holds these models.
A desk box is the other route. A DGX Station keeps inference wherever it sits, in your Canadian office or data centre. You own the hardware, the power and the upkeep, and one box on a desk is not a redundant production server.
XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account or on-premises, sizes the hardware to the model, and builds the workflow on top. Read how we deploy, the Canadian residency guide, and our comparison of on-premise vs client cloud vs managed cloud. If you need a model running with inference kept in Canada, bring one job to a free 30-minute call.
What this sizing leaves out
- No test of our own. Every fit above is arithmetic from file sizes and published memory totals, checked against the engines' recipes. We have not loaded any of the three models.
- Speed. Fitting is not the same as fast. We give no word-pieces-per-second figures, because we measured none.
- Engine overhead. vLLM's commands keep 3% to 10% of GPU memory free, and the engines' own buffers vary by version.
- Two-machine setups. Splitting one model across machines needs fast networking between them. p5 and p5en are built for it, but we did not test the link.
- Prices outside AWS. We confirmed what Azure and Google list in Canada, not what they charge.
For how the three models compare on quality and licence, see Kimi K3.1 Pro vs Kimi K3, MiniMax M3.1 Flash vs DeepSeek V4.1 Flash and the best open-source coding models of 2026.
FAQ
How much GPU memory does Kimi K3.1 Pro need?
At least 1.56 terabytes, the size of its weights on Hugging Face, plus room for conversations. With 20% headroom that is about 1,874 gigabytes. Three H100 machines (1,920 gigabytes) or two H200 machines (2,256 gigabytes) clear it. SGLang's own layouts use four H100 machines or two H200 machines, and vLLM's recipe starts from eight Blackwell B300 GPUs.
Can Kimi K3.1 Pro run on one 8x H100 machine?
No. One H100 machine has 640 gigabytes of GPU memory and Kimi's weights take 1,561 gigabytes, so it is less than half. Even two machines fall 281 gigabytes short. There is no smaller official version: the published weights are already stored at 4 bits. For a single H100 machine, DeepSeek or MiniMax fits.
Can you run DeepSeek V4.1 Flash on a DGX Station?
Yes, with a setup vLLM documents. Its 476 GiB of weights don't fit in the Station's 252 gigabytes of GPU memory, so vLLM keeps the Engram lookup tables and 60 GiB of experts in the Station's slower processor memory. It runs one GPU and holds about 3.09M tokens of working memory, per vLLM's recipe. We have not run it ourselves.
Can a DGX Spark run any of these models?
No. A DGX Spark has 128 gigabytes of memory. The smallest of the three, MiniMax, has 252 gigabytes of weights, and DeepSeek 510 gigabytes. NVIDIA rates the Spark for inference on models up to about 200 billion parameters. For these three you need H100-class machines or larger.
What is the KV cache, and does it change the GPU count?
The KV cache is the model's working memory for a conversation: notes on every word-piece it has read. It grows with conversation length and with how many run at once. For DeepSeek it is tiny, 890 bytes per word-piece. For MiniMax, by our estimate, it is about 61 gigabytes for a single million-token conversation, enough to change how many agents one machine serves.
Sources
- Moonshot AI, Kimi K3.1 Pro model card and files, Hugging Face, weights public by 5 Nov 2026, file sizes read 11 Nov 2026
- DeepSeek, DeepSeek-V4.1-Flash model card and files, Hugging Face, read 11 Nov 2026; KV cache of 890 bytes per token
- MiniMax, MiniMax-M3.1-Flash model card and files and config.json, Hugging Face, read 11 Nov 2026
- vLLM, recipes for Kimi K3.1 Pro, DeepSeek-V4.1-Flash and MiniMax M3.1 Flash, read 11 Nov 2026
- SGLang, cookbook for Kimi K3.1 Pro and MiniMax M3.1, read 11 Nov 2026
- NVIDIA, H100, H200 and A100 product pages, DGX Station and DGX Spark pages, read 11 Nov 2026
- AWS, P4 instances, P5 instances and on-demand pricing; AWS's price list read 11 Nov 2026
- Microsoft, ND H100 v5 sizes, read 11 Nov 2026
- Google Cloud, GPU regions and zones, updated 10 Nov 2026, and accelerator-optimized machines, read 11 Nov 2026
Written by
Cho Yin Yong
Principal AI Solutions Engineer, XY Space
Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.
More from Cho Yin YongShare this article
Work with us
We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.
Start a project