Kimi K3 has 2.8 trillion parameters, and its files take 1,560.9 GB even in the compressed form its maker releases. That is more than any eight-GPU machine in our price records can hold. Running it yourself takes one of NVIDIA's largest B300 servers, or several machines working together.
DeepSeek's models are within easier reach. DeepSeek V4 Flash fits on four 80 GB cards, and V4.1 Flash on one machine with eight. For most people an API is still the practical route, and Novita, a provider we cover, lists both families. The tables below match each model to the smallest machine that holds it, with this week's prices.
Who this is for
- you want to run Kimi K3 or a DeepSeek V4 model on hardware you control
- you need to know which GPU machine holds each model and what it costs by the hour
- you are weighing a rented machine against an API for these models
Not for you if you want a model that runs on one GPU: start with where to run an open-source LLM.
Key numbers
- 2.8 trillion parameters in Kimi K3, according to its Hugging Face page Official page,
- 1,560.9 GB Kimi K3's files as released, in its 4-bit MXFP4 format Official page,
- 510.3 GB the files of DeepSeek V4.1 Flash, small enough for one machine with eight 80 GB cards Official page,
- 900 GB/s NVLink between H100 SXM cards, against 128 GB/s over PCIe Gen5, by NVIDIA's figures Official page,
How big these models are
Each model is a mixture of experts: only part of it works on each token. All of its parameters must still sit in GPU memory, so the size of the files decides the hardware, not the active part. Both makers publish their files already compressed, so these sizes are the smallest form they release.
Kimi K3's files are three times the size of DeepSeek V4.1 Flash's and more than nine times those of V4 Flash.
| Model (maker) | Added to Hugging Face | Parameters | Files as released (GB) | Format | Licence |
|---|---|---|---|---|---|
| Kimi K3 (Moonshot AI) | 13 Jun 2026, updated 2 Sep 2026 | 2.8 trillion, 104 billion active | 1,560.9 | 4-bit (MXFP4) weights | Kimi K3 License |
| DeepSeek V4 Pro 0813 (DeepSeek) | 13 Aug 2026 | about 1.6 trillion, 49 billion active (V4 Pro's model card) | 892.7 | 4-bit (FP4) experts, 8-bit (FP8) for most other weights | MIT |
| DeepSeek V4.1 Flash 1017 (DeepSeek) | 17 Oct 2026 | 763 billion (Hugging Face's count; no active figure published) | 510.3 | 8-bit (FP8), as its page is tagged | MIT |
| DeepSeek V4 Flash 0731 (DeepSeek) | 31 Jul 2026 | 284 billion, 13 billion active (V4 Flash's model card) | 166.9 | 4-bit (FP4) experts, 8-bit (FP8) for most other weights | MIT |
DeepSeek updates its models often, and each version keeps its own repository. V4.1 Flash is the newest Flash model and, at 510.3 GB, a much larger one than V4 Flash. Its page calls the 1017 version a minor update of V4.1 Flash, and the model reads images as well as text.
What hardware each model needs
We keep a tenth of every machine's GPU memory free for conversations and working space, as in our GPU memory guide. Matched against the files, that gives the smallest machine for each model:
DeepSeek V4 Flash fits a four-card H100 server and V4.1 Flash an eight-card one. V4 Pro and Kimi K3 need machines not yet in our records.
| Model | Files (GB) | Smallest machine that holds it | In our price records |
|---|---|---|---|
| DeepSeek V4 Flash 0731 | 166.9 | four 80 GB cards (320 GB) | Unihost LWN-G17 with four H100 cards: €4,399.90 a month (checked 25 October 2026) |
| DeepSeek V4.1 Flash 1017 | 510.3 | eight 80 GB cards (640 GB) | DigitalOcean, eight H100: $35.28 an hour (about $25,754 a month; checked 25 October 2026); UpCloud, eight H100 SXM: €17.36 an hour (about €11,666 a month, with a 28-day monthly cap; checked 25 October 2026); Unihost AZM-12, eight H100 NVLink: €13,890.90 a month (checked 25 October 2026) |
| DeepSeek V4 Pro 0813 | 892.7 | eight H200 cards (1,128 GB), or two machines with eight 80 GB cards | none yet; DigitalOcean lists GPU Droplets with eight H200 cards |
| Kimi K3 | 1,560.9 | eight B300 cards (2.1 TB), or three machines with eight 80 GB cards | none yet; UpCloud lists servers with eight B300 cards |
Monthly costs for hourly prices assume the server runs all month (730 hours), or less where a provider caps the monthly bill.
Two machines can share one model, as the vLLM documentation explains. The model is split across the GPUs inside each machine, and its layers across the machines. That works, but it needs a fast network between the machines and more set-up, so one large machine is simpler when you can rent it.
Why the links between GPUs matter at this size
When a model is split across eight GPUs, the cards exchange data for every token they produce. NVIDIA gives the H100 SXM 900 GB/s over NVLink, its direct link between cards, against 128 GB/s over PCIe Gen5. Check how the cards of an eight-GPU machine are linked before you rent it.
Unihost names its AZM-12 server "8x NVIDIA H100 NVLink", and UpCloud's eight-GPU server uses H100 SXM cards, the version NVIDIA specifies with NVLink.
What that hardware costs
An eight-GPU machine costs the same for every hour it runs, busy or idle. Downloading 510 GB of files to it takes time you pay for as well. Here are the machines in our price records that hold DeepSeek's Flash models:
| Plan | Price | New-customer offer | Link |
|---|---|---|---|
| DigitalOcean GPU Droplet, 8 x NVIDIA H100 8× 80 GB GPUs | $35.28/hour (about $25,754/month) on demand, billed per second; price for the whole 8-GPU machine · VAT not stated checked 25 Oct 2026 | New accounts get a $5 credit that lasts 90 days and covers every product except SaaS add-ons checked 29 Oct 2026 | Visit |
| UpCloud GPU Server, 8 x NVIDIA H100, 96 cores, 1920 GB 8× 80 GB GPUs | €17.36/hour (about €11,666/month; no charge after 28 days of use)30-day month: the last 2 days are free.31-day month: the last 3 days are free.February: no free days (one in a leap year).Started mid-month: it may never reach the cap that month. on demand, billed by the starting hour; price for the whole 8-GPU server checked 25 Oct 2026 | No offer | Visit |
| Unihost AZM-12, 8x NVIDIA H100 NVLink 8× 80 GB GPUs | €13,890.90/month monthly, paid in advance; minimum one month checked 25 Oct 2026 | No offer | Visit |
| Unihost LWN-G17, 4x Nvidia H100 4× 80 GB GPUs | €4,399.90/month monthly, paid in advance; minimum one month; plus a one-off set-up fee of €36.90 checked 25 Oct 2026 | No offer | Visit |
Prices for the whole machine. UpCloud and Unihost prices are in euros and DigitalOcean's in US dollars; we do not convert them.
A working day of tests on DeepSeek V4.1 Flash, eight hours on eight H100 cards, adds up as follows:
| What it is for | Plan and price | Amount | Cost | Link |
|---|---|---|---|---|
| Eight H100 cards, 8 hours | DigitalOcean GPU Droplet, 8 x NVIDIA H100 $35.28/hour (about $25,754/month) on demand, billed per second; price for the whole 8-GPU machine · VAT not stated checked 25 Oct 2026 | 8 hours | $282.24 | Visit |
| Total | $282.24 | |||
The eight hours are our example, not a measurement. DigitalOcean bills by the second.
| What it is for | Plan and price | Amount | Cost | Link |
|---|---|---|---|---|
| Eight H100 SXM cards, 8 hours | UpCloud GPU Server, 8 x NVIDIA H100, 96 cores, 1920 GB €17.36/hour (about €11,666/month; no charge after 28 days of use)30-day month: the last 2 days are free.31-day month: the last 3 days are free.February: no free days (one in a leap year).Started mid-month: it may never reach the cap that month. on demand, billed by the starting hour; price for the whole 8-GPU server checked 25 Oct 2026 | 8 hours | €138.88 | Visit |
| Total | €138.88 | |||
The eight hours are our example, not a measurement. UpCloud bills each started hour.
Run all month, the same machines cost what the boxes above show as a monthly figure. The question then becomes whether your traffic would keep them busy. Our guide to API, rented GPU or your own hardware works through it.
Through an API instead
An API runs the model on the provider's machines and charges per million tokens, so the size of the model is no longer your problem. On 26 October 2026, Novita AI's model page listed Kimi K3 and Kimi K3.1 Pro. It also listed DeepSeek V4 Pro, V4 Flash 0731 and V4.1 Flash, each priced per million tokens. Its prices are not in our weekly records yet, so we do not show them here.
Novita's terms describe a zero data retention policy, with exceptions for legal duties, support and automated safety checks. Its privacy policy mentions transfers to the US, and its legal pages do not name the company behind it. Our guide to AI hosting in Europe compares the data terms of every provider we cover.
If your data must stay in Europe, Infomaniak serves the previous Kimi model, Kimi K2.6, from its own data centres in Switzerland. It does not offer Kimi K3 or DeepSeek V4.
Is it safe to run DeepSeek or Kimi yourself?
Running the weights on a server you rent is not the same as using the makers' own apps. On your server, the model is a set of files that your serving software loads. Your prompts go only where your own software sends them.
The files themselves use safetensors, a format Hugging Face created as a safe alternative to pickle. The Python documentation warns that loading pickle data can run code, which is why the newer format matters.
Code is a separate question. Kimi K3's page shows the Transformers library loading it with trust_remote_code=True, which runs Python code from the model's repository. Read that code before you run it, or use vLLM or SGLang, which the page lists among the supported engines.
The licences differ too. DeepSeek publishes its models under the MIT licence. Kimi K3 comes under Moonshot AI's own Kimi K3 License, so read its terms before you build a product on it. In our view, also test any model on your own tasks before you rely on its answers, wherever it runs.
Smaller open models that run on one card
Most people who want their own model need far less than this. OpenAI says its gpt-oss-120b runs on a single 80 GB card, and many capable models fit 24 GB or 32 GB. These guides cover them:
- How much GPU memory an LLM needs, model by model
- Where to run an open-source LLM, from a CPU-only server to eight GPUs
- gpt-oss-120b: what it needs and where to run it
- Running Qwen3.8-27B on one 24 GB or 32 GB GPU
Mistakes to avoid
Sizing a mixture-of-experts model by its active parameters
Instead: Size it by its files: Kimi K3 has 104 billion active parameters but 1,560.9 GB to load.
Filling a machine's memory with the model alone
Instead: Keep about a tenth free for conversations and working space.
Renting eight GPUs without checking how they are linked
Instead: Look for NVLink or the SXM version of the card.
Running a model's own code without reading it
Instead: Read the repository's code, or use an engine that supports the model.
Leaving an hourly eight-GPU machine running overnight
Instead: Delete it when the job is done: it bills every hour it exists.
Other questions
How much VRAM does Kimi K3 need?
Its files alone take 1,560.9 GB. With a tenth of the memory kept free, that means about 1,734 GB of GPU memory. That is more than the 1,128 GB of eight H200 cards but within the 2.1 TB of an eight-card B300 machine.
Can I run DeepSeek V4 Flash on my own PC?
Not on one consumer card: its files take 166.9 GB, and the largest consumer card we list, the RTX 5090, has 32 GB. Four 80 GB data-centre cards hold it.
Is DeepSeek safe to run locally?
On your own server, the model runs where you put it and sends your prompts nowhere unless your software does. Check any code you run alongside it, read its licence and test its answers on your own tasks.
Is an API cheaper than renting the hardware?
For most traffic, yes, because an eight-GPU machine costs the same whether it is busy or idle. Our guide to API, rented GPU or your own hardware shows how to find your break-even point.
How we worked out these figures
Parameter counts and formats come from each model's Hugging Face page, and file sizes from its repository's file list, read on 26 October 2026. We count 1 GB as a billion bytes. Machine sizes come from NVIDIA's product pages. The smallest machine for each model keeps a tenth of its GPU memory free, as in our GPU memory guide.
We ran nothing for this page. Prices come from our weekly price checks, each with its own date. We plan to read these pages again every three months, and sooner when a maker releases a new version.
Sources and test notes
- Kimi K3 on Hugging Face, read
- DeepSeek V4 Pro 0813 on Hugging Face, read
- DeepSeek V4 Pro on Hugging Face (parameters and format), read
- DeepSeek V4.1 Flash 1017 on Hugging Face, read
- DeepSeek V4 Flash 0731 on Hugging Face, read
- DeepSeek V4 Flash on Hugging Face (parameters and format), read
- NVIDIA H100 specifications (memory and NVLink), read
- NVIDIA H200 specifications, read
- NVIDIA HGX platform (B200 and B300 memory), read
- vLLM documentation: parallelism and scaling, read
- UpCloud pricing (eight-GPU servers), read
- DigitalOcean GPU Droplet pricing, read
- Novita AI model library, read
- Novita AI terms of service, read
- Novita AI privacy policy, read
- Infomaniak AI Services prices (models served), read
- Hugging Face safetensors documentation, read
- Python documentation: pickle (security warning), read
- OpenAI gpt-oss-120b on Hugging Face (one 80 GB GPU), read
We ran no test for this page. We read the model facts and file sizes on the makers' Hugging Face pages on 26 October 2026. NVIDIA's figures, the vLLM documentation and the providers' machine lists were read on 25 and 26 October 2026. Every price comes from our weekly price records and shows its own date.