Kimi K2.5
Kimi K2.5 needs roughly 683.2 GB VRAM at Q4_K_M quantization (2241.2 GB at FP16). 1 GPUs we track can run it fully in VRAM at 8k context.
1 GPU runs this natively · 0 with CPU offload
Kimi K2.5 is a Mixture of Experts (MoE) large language model with 1000B total parameters but only 32B active per token developed by Moonshot AI. Released in January 2026, it supports text and vision and video inputs with a 256K context window, released under the Modified MIT license, allowing commercial use. Native multimodal MoE with 384 routed experts, 8 selected plus one shared expert, a 400M-parameter MoonViT encoder, MLA, thinking/instant modes, tool use, and agent-swarm workflows. Scores are Moonshot-reported in thinking mode; LiveCodeBench is v6. The released weights use native INT4, while Ollama's official kimi-k2.5 tag is cloud-only and must not be treated as a local hardware path.
To run Kimi K2.5 locally, you need approximately 683.2 GB of VRAM at Q4_K_M quantization with 8k context. 1 of the GPUs we track can run it fully in VRAM, with a further 0 able to offload to system RAM. At Q4_K_M it requires 683.2 GB — more than any single consumer GPU. A multi-GPU server or 80 GB datacenter GPU is required. At Q8_K_M (1191.7 GB), you get near-FP16 quality while still fitting on large server or multi-GPU setups. FP16 requires 2241.1 GB, limiting it to datacenter-class hardware with 80 GB+ VRAM. As a MoE model, inference speed depends on active parameters (32B) rather than total size, so once it fits in VRAM it runs noticeably faster than dense models of the same parameter count.
Its MMLU-Pro score of 87.1 places it among the strongest open-weight models available. Its LiveCodeBench score of 85.0 makes it a solid choice for code generation and completion. The license allows commercial use.
VRAM at each quantization
Numbers here are computed at 8k context. Because KV cache grows linearly with context length, expect higher totals at longer sequence lengths.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 4000.0 GB | 1.02 GB | 4481.1 GB |
| BF16 | 2000.0 GB | 1.02 GB | 2241.2 GB |
| FP16 | 2000.0 GB | 1.02 GB | 2241.2 GB |
| Q8_0 | 1063.0 GB | 1.02 GB | 1191.7 GB |
| Q6_K | 821.0 GB | 1.02 GB | 920.7 GB |
| Q5_K_M | 712.0 GB | 1.02 GB | 798.6 GB |
| Q4_K_Mrec | 609.0 GB | 1.02 GB | 683.2 GB |
| Q3_K_M | 481.0 GB | 1.02 GB | 539.9 GB |
| Q2_K | 381.0 GB | 1.02 GB | 427.9 GB |
| NVFP4cuda | 500.0 GB | 1.02 GB | 561.1 GB |
KV cache is calculated at 8k context (FP16). Note that NVFP4 only runs on CUDA GPUs. Turn on TurboQuant in the calculator above for lower KV cache estimates.
Benchmarks
GPUs that run Kimi K2.5 natively (1)
- Apple M3 Ultra (512GB)Q2_K · 15.7 t/s
Notes
Native multimodal MoE with 384 routed experts, 8 selected plus one shared expert, a 400M-parameter MoonViT encoder, MLA, thinking/instant modes, tool use, and agent-swarm workflows. Scores are Moonshot-reported in thinking mode; LiveCodeBench is v6. The released weights use native INT4, while Ollama's official kimi-k2.5 tag is cloud-only and must not be treated as a local hardware path.
How to run Kimi K2.5 locally
Q4_K_M needs 683.2 GB — needs multiple datacenter-class GPUs (80 GB+ each).
vLLM (native INT4)
vllm serve moonshotai/Kimi-K2.5 --trust-remote-codeLM Studio: Community GGUF builds exist, but the official release is native INT4 and remains roughly a half-terabyte-class deployment. The official Ollama entry is cloud-only.
Why this quantization? Kimi K2.5 was trained and released with a native INT4 path, so Q4 is the closest calculator representation. It still requires multi-GPU or very large unified-memory hardware; MLA keeps the KV cache comparatively small but does not reduce the need to load the full 1T expert pool.
Who is Kimi K2.5 for?
Organizations self-hosting a frontier multimodal agent model for coding, visual reasoning, tool use, and coordinated agent workflows.
Best for
- Visual coding from screenshots or designs
- Long-horizon tool-using agents
- Image/video analysis with reasoning
- Agent-swarm task decomposition
Not ideal for
- Consumer local inference
- Non-thinking low-latency workloads
- Treating the Ollama cloud tag as a local model
Frequently asked questions
- What are the VRAM requirements for Kimi K2.5?
- Kimi K2.5 requires approximately 683.2 GB of VRAM at Q4_K_M quantization, 1191.7 GB at Q8, and 2241.1 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.
- How many parameters does Kimi K2.5 have?
- Kimi K2.5 has 1000 billion total parameters, but only 32 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
- Is Kimi K2.5 good for coding?
- Yes. Kimi K2.5 scores 85.0 on LiveCodeBench, demonstrating strong code generation and completion capabilities.
- Can Kimi K2.5 run on a 16 GB GPU?
- No. At Q4_K_M, Kimi K2.5 needs 683.2 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
- Can Kimi K2.5 run on a 24 GB GPU?
- No. Even at Q4_K_M, Kimi K2.5 needs 683.2 GB. Consider a multi-GPU server with 80 GB+ total VRAM.
- What is the smallest quantization for Kimi K2.5 that fits in 24 GB of VRAM?
- Kimi K2.5 cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 427.9 GB at Q2_K.
- What GPU do I need to run Kimi K2.5 locally?
- You need a multi-GPU server. At Q4_K_M, Kimi K2.5 needs 683.2 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.