Kimi K2.5

What is Kimi K2.5?

Kimi K2.5 is Moonshot AI's MoE open-weights language model. Released in 2026-01, it ships in 1 size (MoE) and is licensed Modified MIT. The most popular use case is Kimi K2.5 — January 2026 frontier update.

VRAM and hardware

The MoE size fits in 256 GB of VRAM.

Size	Min VRAM (Q4_K_M)	Recommended VRAM	Tokens/sec on 3090
MoE	256 GB	384 GB	~10 tok/s

How to run Kimi K2.5 locally

Option 1: Ollama (simplest)

ollama pull kimi-k2-5
ollama run kimi-k2-5

Option 2: Mullama (production)

mullama pull kimi-k2-5
mullama run kimi-k2-5

Option 3: llama.cpp (CLI)

# Download a GGUF from Hugging Face (search "kimi-k2-5 gguf")
./llama-cli -m kimi-k2-5.Q4_K_M.gguf -p "Hello, AI!"

Option 4: Python with Mullama or llama-cpp-python

from mullama import Model, Context
model = Model.load("kimi-k2-5.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256))

What you can build with Kimi K2.5

• agentic
• long-context
• tool use

Hardware it fits on

• Apple Silicon (8GB): MoE in MLX
• Apple Silicon (16GB): MoE in MLX
• Apple Silicon (32GB): MoE in MLX or GGUF
• Apple Silicon (64GB+): MoE in MLX
• RTX 3090 (24GB): MoE at Q4_K_M
• RTX 4090 (24GB): same as 3090, ~30% faster
• RTX 5090 (32GB): MoE at Q8_0
• 2× RTX 6000 Ada (96GB total): MoE at Q4_K_M

Related models in the Kimi family

• Kimi K2 — 1T (32B active MoE) · frontier agentic with 1T total / 32B active MoE
• Kimi K2.6 — MoE · Kimi K2.6 — March 2026 update
• Kimi K2.7 Code — MoE · Kimi K2.7 code-specialized variant (June 2026)

Other models you might consider

• Llama 3.1 (Meta) — 8B, general-purpose chat, agentic workflows, code
• Llama 3.2 (Meta) — 1B, low-end hardware, mobile, edge deployment
• Llama 3.2 Vision (Meta) — 11B, vision + chat on desktop GPUs
• Llama 3.3 (Meta) — 70B, best-in-class 70B for desktops
• Llama 4 (Meta) — Scout 17B (109B MoE), vision + long context (10M tokens)
• Qwen 2.5 (Alibaba) — 0.5B, strongest 7B-32B on consumer hardware