Kimi K2

What Kimi K2 is for

Kimi K2 is a large mixture-of-experts model: only a fraction of its parameters activate per token, but all of them have to be resident, so total size decides whether you can run it at all. Moonshot AI published it in 2025-07 under Modified MIT with a single 1T (32B active MoE) size and 128K of context. The catalogue records its strength as frontier agentic with 1T total / 32B active MoE.

Sizes and memory

Our catalogue does not carry a per-size memory breakdown for Kimi K2. What it records instead is a floor of 384 GB and a recommended 768 GB for the default configuration. Until a per-size figure exists, treat those two numbers as the whole guidance rather than reading a size table that would have to be invented.

Hardware that runs Kimi K2

Without a per-size memory figure there is no honest way to say which card holds Kimi K2. Its recorded floor of 384 GB rules out every consumer card in our hardware guide, which tops out well below that.

What people use it for

The catalogue lists 4 recorded uses for Kimi K2: frontier agentic, long-context, tool use, research. It is tagged MoE, tools, agentic, frontier, long-context.

Where Kimi K2 sits in the Kimi family

Our catalogue holds 4 Kimi entries. Ordered by release date, Kimi K2 (2025-07) is the 1st of them, and Kimi K2.7 Code (2026-06) is newer.

It ships under the Modified MIT, which is neither a standard open-source licence nor a closed one. Read the terms before shipping anything commercial on top of it.

Context window: 128K

128K tokens is roughly 96,000 words at the usual ratio, which is a small repository or a whole book. Two cautions come with a window this size: the KV cache at full length can rival the weights for memory, and models rarely use the far end of their advertised context as well as they use the near end. Treat it as headroom, not as a promise.

How to run Kimi K2 locally

Our catalogue does not record registry tags, so look the current tag up in the Ollama library or on Hugging Face before pasting these. The commands below are the shape of the workflow, not a copy-and-paste recipe.

Option 1: Ollama (simplest)

# find the tag for Kimi K2 at ollama.com/library
ollama run <tag>

Option 2: Mullama (production)

mullama pull <tag>
mullama run <tag>

Option 3: llama.cpp (CLI)

# download a GGUF from Hugging Face, searching for "Kimi K2 GGUF"
./llama-cli -m model.Q4_K_M.gguf -p "Hello, AI!"

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

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

Sources

Model record from src/data/models.json (2026-06-29); card capacities from src/data/gpus.json (2026-06-29). Upstream: Curated from ollama.com/library + community benchmarks + paperswithcode + huggingface.