EDGE OSMOONSHOT AI · 1T PARAMETERS
Kimi K2 hardware and VRAM requirements
Moonshot AI’s 1 trillion parameter mixture-of-experts model with 32B active, the predecessor to Kimi K3. Kimi K2 is a mixture-of-experts model with 32B of its 1T parameters work on each token. At Block-FP8 it needs about 1.05 TB of GPU memory for one user, and about 1.06 TB to serve a team of 16 with 8K context each.
Total parameters
1T
Active per token
32B
Context length
128K tokens
Released as
Block-FP8 (native)
Minimum VRAM
579 GB at 4-bit
For a team of 16
1.06 TB at Block-FP8
VRAM needed to run Kimi K2
GPU memory by precision, number of concurrent users and context length. Each user needs their own KV cache, so memory grows with users and context.
| Precision | Weights only | 1 user, 8K | 16 users, 8K | 16 users, 32K | LoRA fine-tune |
|---|---|---|---|---|---|
| Block-FP8 (native) Released | 1 TB | 1.05 TB | 1.06 TB | 1.09 TB | 1.21 TB |
| 4-bit Community quantization (estimated) | 550 GB | 579 GB | 588 GB | 615 GB | 741 GB |
| BF16 Upcast | 2 TB | 2.1 TB | 2.11 TB | 2.14 TB | 2.26 TB |
Full fine-tuning in mixed precision needs about 16.1 TB: weights, gradients and optimizer states at roughly 16 bytes per parameter. LoRA trains small adapter layers on a frozen copy of the model instead, so it needs far less.
GPUs that can run Kimi K2 on-premise
For a team of 16 using Kimi K2 at Block-FP8 with 8K context each (about 1.06 TB), smallest fit first.
NVIDIA DGX B200
1.44 TB per system · 236 GB to spare.
AMD Instinct MI300X
192 GB per GPU · Split across 8 GPUs in one server, 322 GB to spare.
AMD Instinct MI325X
256 GB per GPU · Split across 8 GPUs in one server, 783 GB to spare.
NVIDIA DGX Spark
128 GB per system · Needs more than 2 linked systems.
NVIDIA H100 NVL
94 GB per GPU · Needs more than one 8-GPU server.
NVIDIA L40S
48 GB per GPU · Needs more than one 8-GPU server.
NVIDIA RTX PRO 6000 Blackwell
96 GB per GPU · Needs more than one 8-GPU server.
NVIDIA H200 NVL
141 GB per GPU · Needs more than one 8-GPU server.
Some of these setups need several GPUs working as one. Buying the GPUs is the easy part. Vanik Edge OS splits the model across them, keeps it running if a server fails and manages it all from one console. See how Edge OS clusters GPU servers
Weighing two of these? Compare AI hardware side by side, for example H100 vs H200, RTX PRO 6000 vs RTX 5090 or DGX Spark vs Mac Studio.
Running Kimi K2 for your team?
We will size the GPU server for your users and data, and set up Kimi K2 on it inside your own building. Free, no obligation.
Kimi K2 at a glance
FROM SIZING TO RUNNING
Sizing the hardware is the easy part.
Getting several GPUs to work as one system for a whole company is where most projects slow down. That is the part Vanik Edge OS does for you.
STEP 01 · THIS TOOL
Size it
Work out the GPU memory your models need and which GPUs or servers can hold them.
STEP 02 · VANIK EDGE OS
Cluster it
Edge OS joins your GPU servers into one cluster, splits a large model across them and keeps it running if a server fails. NVIDIA and AMD servers work side by side.
STEP 03 · VANIK EDGE OS
Run it
Your team gets AI apps with single sign-on and an audit trail, managed from one console, with your data staying inside your building.
Kimi K2 GPU questions
What is Kimi K2?
Kimi K2 is Moonshot AI’s open-weight mixture-of-experts model with 1 trillion parameters, of which 32B work on each token. It has a 128K-token context and was the predecessor to Kimi K3.
Is Kimi K2 open source?
Yes. Moonshot AI publishes Kimi K2’s weights under a modified MIT license, so you can download and self-host it.
Is Kimi K2 free?
The weights are free to download under the modified MIT license. Self-hosting it still needs about a terabyte of GPU memory.
How to run Kimi K2 locally?
To run Kimi K2 locally you need about 579 GB of GPU memory at 4-bit for one user (1.06 TB for a team of 16), system RAM of at least 550 GB so the weights can be loaded, around 550 GB of fast NVMe storage for the model files, and an inference engine such as vLLM or SGLang. A team of 16 fits on NVIDIA DGX B200.
How big is Kimi K2?
Kimi K2 has 1T parameters in total, of which 32B work on each token. The weights take about 550 GB at 4-bit.
How much VRAM to run Kimi K2?
Kimi K2 needs about 1.05 TB of GPU memory at Block-FP8 for one user with 8K context: 1 TB for the weights plus the KV cache and runtime overhead. Serving 16 people at once with 8K context each needs about 1.06 TB.
How many GPUs does Kimi K2 need?
For 16 users with 8K context at Block-FP8 it needs about 1.06 TB. That fits on NVIDIA DGX B200.
Why does Kimi K2 need so much memory if only 32B parameters are active?
It is a mixture-of-experts model. Only 32B parameters work on each token, which keeps it fast, but a different set of experts is picked for every token, so all 1T parameters must stay loaded in GPU memory.
Can I fine-tune Kimi K2?
LoRA fine-tuning at Block-FP8 needs roughly 1.21 TB with 4K sequences; it trains small adapter layers on a frozen copy of the model. Full fine-tuning needs about 16.1 TB, which calls for a multi-GPU training cluster.
How are these numbers calculated?
Weights are parameters × bytes per parameter. The KV cache uses 68.6 KB per token per user. We add about 5% of the weights plus 1 GB for runtime overhead and keep 10% of each GPU free. Figures are planning estimates; real usage depends on the inference engine and settings.
VRAM for similar models
Specifications from GitHub: MoonshotAI/Kimi-K2. Memory and bandwidth are manufacturer figures; GPU counts are rounded to 1, 2, 4 or 8 per server. All figures are planning estimates. Last checked November 10, 2026.
Run Kimi K2 inside your own building
Vanik Edge OS runs open-weight models on NVIDIA DGX, AMD Instinct and other GPU servers on your premises, so your data never leaves.