What GPU do I need to run moonshotai/Kimi-K2.5?
A 1T-parameter (32B active) native multimodal agentic model. 1026.9B parameters, published in Native INT4. View on Hugging Face
What Kimi K2.5 is
Kimi K2.5 is Moonshot AI's native multimodal agentic model, a 1T-parameter mixture-of-experts model (32B active) continually pretrained on approximately 15 trillion mixed vision-language tokens atop Kimi-K2-Base. It integrates vision and language understanding with agentic tool use, instant and thinking modes, and an "Agent Swarm" mode that decomposes tasks across dynamically instantiated sub-agents. Its weights publish in Moonshot's native INT4 quantization, the same scheme as Kimi-K2-Thinking.
License note: MIT terms, with one modification: a commercial product or service with more than 100 million monthly active users or more than $20 million/month revenue must prominently display "Kimi K2.5" in its user interface. Facts in this section are sourced from Kimi K2.5's Hugging Face model card, not benchmarked by Aquanode.
What it's used for
- Vision-language agentic tasks
- Coding from visual specifications (UI designs)
- Multi-agent tool-use workflows
Benchmarks (published by Moonshot AI)
Published by Moonshot AI, not measured by Aquanode.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision, plus a fixed overhead for KV-cache, activations, and fragmentation. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| Native INT4 (native) | 554.3 GB | 1128.0 GB | No capable live offer found | – | – |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
Moonshot's own deployment guide gives an example vLLM/SGLang command serving Kimi-K2.5 on a single H200 node at tensor-parallel-8; NVIDIA's own H200 spec lists 141GB HBM3e per GPU, so 8×141GB ≈ 1128GB total.
How to run Kimi K2.5
Run Kimi K2.5 with vLLM
From moonshotai/Kimi-K2.5's own deployment guide: an example serving it on a single H200 node at tensor-parallel-8.
vllm serve $MODEL_PATH -tp 8 --mm-encoder-tp-mode data --trust-remote-code --tool-call-parser kimi_k2 --reasoning-parser kimi_k2Source: https://huggingface.co/moonshotai/Kimi-K2.5/raw/main/docs/deploy_guidance.md
Run Kimi K2.5 with SGLang
From moonshotai/Kimi-K2.5's own deployment guide, same single H200 node at tensor-parallel-8.
sglang serve --model-path $MODEL_PATH --tp 8 --trust-remote-code --tool-call-parser kimi_k2 --reasoning-parser kimi_k2Source: https://huggingface.co/moonshotai/Kimi-K2.5/raw/main/docs/deploy_guidance.md
Deploy Kimi K2.5 on Aquanode
Aquanode has no one-click deploy template for Kimi K2.5; you install the inference engine yourself with the commands below. Aquanode sells GPU pods billed per second, not a hosted inference API.
- Launch a bare GPU pod sized to the requirement above (1128 GB VRAM or more).
- Open a terminal on the pod, or save one of the commands above as a startup script so it runs automatically the first time the pod boots.
- Run the command and connect to the resulting endpoint.
Launch a GPU podWeight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Kimi K2 models
- Kimi-K2-Instruct-0905 (1026.5B, F8_E4M3)
- Kimi-K2-Base (1026.5B, F8_E4M3)
- Kimi-K2-Instruct (1026.4B, F8_E4M3)
- kimi-k2.6-eagle3-mla (3.0B, BF16)
Related reading: H100 pricing and specs, The best GPUs for AI, ranked, and Best GPU for LLM inference.
