Short answer
If by “Kimi K2 5” you mean Kimi K2.5, its requirements depend on how you plan to use it. For one documented local setup using KTransformers and SGLang, the guide lists five categories: GPU, CPU, system memory, model-weight storage, and software. Those are requirements for that third-party deployment path—not universal Kimi K2.5 requirements.
For hosted production API use, there is an important current limitation: the official Kimi API documentation says Kimi K2.5 was retired across the platform on August 31, 2026. The same documentation points new users to moonshot/kimi-k3.1-pro. So the local checklist below may help evaluate self-hosting, but it is not a checklist for a currently available official hosted K2.5 API. (platform.kimi.ai)
Five requirements in one local deployment guide
The figures below come from KTransformers’ tutorial for running Kimi K2.5 with SGLang and KT-Kernel, using CPU–GPU heterogeneous inference. The guide describes offloading model experts to the CPU. It is implementation-specific documentation, not a Moonshot-issued hardware standard. (github.com)
| Category | Requirement stated by the guide | Qualification |
|---|---|---|
| 1. GPU / VRAM | NVIDIA “RTX 2x4090 48GB,” or an equivalent setup with at least 48 GB total VRAM | The guide separately gives a 4× RTX 4090 server command as an example. That example is not the same as its stated minimum. |
| 2. CPU | x86 CPU with AVX512F support; Intel Sapphire Rapids is given as an example | This is for the documented KTransformers/SGLang setup. |
| 3. System RAM | At least 600 GB | Do not merge this number with memory estimates from other runtimes or community posts. |
| 4. Storage | About 600 GB for model weights | The guide specifies native INT4 weights and the same weight folder for CPU and GPU. |
| 5. Software and runtime | KT-Kernel, the kvcache-ai fork of SGLang, a CUDA toolkit compatible with the GPU, and Hugging Face CLI | CUDA 12.8+ is recommended, not stated as an absolute minimum. The guide says a cuDNN reinstall may be needed and gives pip install nvidia-cudnn-cu12==9.16.0.29 as an example. |
The guide’s example launch command uses four RTX 4090s and --tensor-parallel-size 4; it should not be presented as the minimum GPU configuration. Also, KTransformers has other Kimi K2.5 deployment documentation with different assumptions, reinforcing that these figures apply to a particular path rather than to every local installation. (github.com)
What the local API example does—and doesn’t—mean
After launching its local server, the tutorial demonstrates an OpenAI-compatible request to http://localhost:31245/v1/chat/completions using the model name Kimi-K2.5. That is a request to your locally running SGLang server. It does not establish a hosted Moonshot endpoint, hosted service limits, or current official API availability for K2.5. (github.com)
If you want a production API
The official API documentation’s retirement notice means you should not plan a new integration around Kimi K2.5 on the official hosted platform. It recommends moonshot/kimi-k3.1-pro for new users. Before implementing, use the current Kimi API model list and quickstart to confirm the selected model’s identifier, endpoint, authentication, supported request formats, and limits. Those API details are model- and platform-specific; the local hardware table above does not answer them. (platform.kimi.ai)
This retirement notice concerns the official Kimi API platform. It does not by itself establish whether a separate third-party provider offers K2.5; check that provider’s current model catalog and documentation if you are considering one.
Choose by workload, not just by model name
Moonshot’s K2.5 product page describes text, image, and video input; visual-to-code workflows; and agent-generated work such as documents, slides, spreadsheets, websites, and research reports. For Web and App use, it describes Instant for quick questions, Thinking for deeper or step-by-step reasoning, Agent for research and structured content creation, and Agent Swarm (Beta) for larger tasks that can be split and run in parallel. These are product descriptions—not independent performance guarantees—and they do not establish that the same modes or features are available through every API provider. (kimi.ai)
A practical choice:
- You need the official hosted API: don’t assume K2.5 is still selectable; its retirement date has passed. Evaluate the currently listed model, such as the recommended K3.1 Pro, against your task and API needs.
- You need to self-host K2.5: first choose a specific runtime and weight format, then follow the matching deployment guide. Use the five-category table only as a checklist for the cited KTransformers/SGLang path.
- You are considering another provider: verify that provider’s K2.5 availability, model ID, authentication, supported modalities, limits, and service terms directly. OpenAI-compatible request syntax alone does not prove that providers offer the same model or capabilities.
Bottom line: for the cited local setup, plan for at least 48 GB total VRAM, AVX512F-capable x86, 600 GB system RAM, roughly 600 GB of weight storage, and its specified software stack. For official hosted API selection, K2.5 is retired; check the current model catalog rather than treating the local checklist as an API requirement.
Sources
Use Kimi in OpenClaw — Kimi API Platform
Model List — Kimi API Platform
Quickstart — Kimi API Platform
Running Kimi-K2.5 with SGLang and KT-Kernel — KTransformers