Kimi Moonshot AI Released 2025-07 MoEtoolsagentic

Kimi K2

Moonshot AI's Kimi K2. 1 trillion total parameters, only 32B active per token. Frontier-class agentic capability with strong tool use. One of the most-pulled models in the Ollama library in 2026.

Best for frontier agentic with 1T total / 32B active MoE
Sizes 1T (32B active MoE)
Context 128K
License Modified MIT
Min VRAM (default size, Q4) 384 GB
Rec VRAM 768 GB

What is Kimi K2?

Kimi K2 is Moonshot AI's MoE open-weights language model. Released in 2025-07, it ships in 1 size (1T (32B active MoE)) and is licensed Modified MIT. The most popular use case is frontier agentic with 1T total / 32B active MoE.

VRAM and hardware

The 1T (32B active MoE) size fits in 384 GB of VRAM.

Size Min VRAM (Q4_K_M) Recommended VRAM Tokens/sec on 3090
1T (32B active MoE) 384 GB 576 GB ~9 tok/s

How to run Kimi K2 locally

Option 1: Ollama (simplest)

ollama pull kimi-k2
ollama run kimi-k2

Option 2: Mullama (production)

mullama pull kimi-k2
mullama run kimi-k2

Option 3: llama.cpp (CLI)

# Download a GGUF from Hugging Face (search "kimi-k2 gguf")
./llama-cli -m kimi-k2.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.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

  • frontier agentic
  • long-context
  • tool use
  • research

Hardware it fits on

  • Apple Silicon (8GB): 1T (32B active MoE) in MLX
  • Apple Silicon (16GB): 1T (32B active MoE) in MLX
  • Apple Silicon (32GB): 1T (32B active MoE) in MLX or GGUF
  • Apple Silicon (64GB+): 1T (32B active MoE) in MLX
  • RTX 3090 (24GB): 1T (32B active MoE) at Q4_K_M
  • RTX 4090 (24GB): same as 3090, ~30% faster
  • RTX 5090 (32GB): 1T (32B active MoE) at Q8_0
  • 2× RTX 6000 Ada (96GB total): 1T (32B active MoE) at Q4_K_M

Related models in the Kimi family

  • Kimi K2.5 — MoE · Kimi K2.5 — January 2026 frontier update
  • 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