How to Run MiniMax M3 on Ollama: Cloud Setup Guide (2026)

What MiniMax M3 Is and Why Ollama Runs It in the Cloud

MiniMax M3 is an open-weight large language model from MiniMax, a Shanghai-based AI lab. The model uses a mixture-of-experts architecture with 428 billion total parameters and roughly 23 billion active per token, built around MiniMax Sparse Attention (MSA), a mechanism that uses a lightweight index branch to scan incoming tokens and decide which blocks of past tokens are worth attending to, on top of a Grouped Query Attention base. MiniMax trained M3 on mixed text, image, and video data from the very first pretraining step, rather than adding multimodal support after the fact, across a training corpus on the order of 100 trillion tokens. The weights live on Hugging Face at `MiniMaxAI/MiniMax-M3` under the MiniMax Community License: free for personal use, self-hosted experimentation, and non-commercial research, with commercial use requiring written authorization from MiniMax. The model targets agentic coding, long-running tool use, and native image and video understanding in one checkpoint.

The reason MiniMax M3 only shows up on Ollama as a cloud model comes down to the same math that applies to Kimi K2.6 and GLM 5.2. Even at the lowest practical quantization, Unsloth's dynamic GGUF build needs around 140 GB of combined RAM and VRAM, which rules out almost every desktop and single-GPU workstation. Ollama's `:cloud` tag sidesteps that entirely. `ollama run minimax-m3:cloud` sends your prompt to MiniMax's infrastructure through Ollama's servers and streams the response back, using the same command syntax as a model that actually lives on your disk.

Here's how the MiniMax line compares on Ollama as of June 2026:

For most people searching "MiniMax M3 Ollama" today, `minimax-m3:cloud` is the only option in Ollama's official library. On MiniMax's own benchmark numbers, M3 scores 59.0% on SWE-Bench Pro and 83.5 on BrowseComp, ahead of Claude Opus 4.7's 79.3 on that same agentic browsing benchmark, though Opus and a few other frontier closed models still lead on raw coding scores.

Set Up Ollama Cloud and Run MiniMax M3

Running MiniMax M3 through Ollama takes three steps: install Ollama, sign in, and run the model. Nothing here downloads a multi-hundred-gigabyte file. The whole setup takes under five minutes on any machine with a working internet connection.

Step 1: Install Ollama

# Linux and macOS, one-command installer
curl -fsSL https://ollama.com/install.sh | sh

On Windows, download the installer from ollama.com/download, or use winget:

winget install Ollama.Ollama

Verify the installation:

ollama --version
# Expected: ollama version 0.12.x or higher

Step 2: Sign In to Ollama Cloud

ollama signin

This prints a sign-in URL and opens your browser. Create a free account at ollama.com, or log in if you already have one, then approve the device. The terminal confirms with a message similar to:

Signing in to ollama.com...
Signed in as your-username

Step 3: Run MiniMax M3 from the Terminal

ollama run minimax-m3:cloud

Ollama fetches a small manifest, a few KB rather than the model weights, since inference happens remotely, then drops you into a prompt:

pulling manifest
pulling 9b1f4a2e... 100% ▕████████████████▏  4.0 KB
success
>>> Send a message (/? for help)

Type a prompt to test it:

>>> Write a Python function that paginates a large Postgres query without loading the full result set into memory.

The first response after signing in can take 10-30 seconds while Ollama establishes the cloud session. After that, responses stream back at normal speed.

Step 4: Verify the Model

ollama list

`minimax-m3:cloud` appears in the list at a few KB rather than hundreds of gigabytes. That's expected: this is a cloud passthrough entry, not a downloaded model.

Switching Between Local and Cloud Models

`ollama run` works the same way for local and cloud models, so you can keep both on one machine. Pull a small local model alongside MiniMax M3:

ollama pull gemma3:4b

`ollama list` now shows both `gemma3:4b` (a multi-gigabyte local download) and `minimax-m3:cloud` (a manifest-only cloud entry). Switch between them by changing the model name in `ollama run` or in your application's API request. Use the small local model for routine tasks, and save MiniMax M3 for the harder, longer-running jobs where the bigger context window and coding benchmarks actually pay off.

Running MiniMax M3 Locally Instead of in the Cloud

If you have the hardware, MiniMax M3 can run entirely on your own machine, just not through Ollama's official library tag. Unsloth publishes dynamic GGUF quantizations of the model that work with llama.cpp, and Ollama can pull a GGUF repo directly from Hugging Face even when that model isn't in Ollama's own library.

Memory requirements by quantization

The full, unquantized model is closer to 850 GB at 16-bit precision. The 1-bit and 2-bit dynamic quants are the practical starting point for anyone testing the local route, and both still need a serious multi-GPU rig or a high-memory server rather than a single consumer card.

Pull the GGUF directly through Ollama

ollama run hf.co/unsloth/MiniMax-M3-GGUF:UD-Q4_K_M

This pulls Unsloth's quantized weights straight from Hugging Face and runs them through Ollama's local engine, no separate download tool needed. Expect the pull itself to take a while given the file sizes in the table above.

For full MSA support and the real speed benefit at long context, MiniMax's own model card recommends SGLang or vLLM over llama.cpp, both of which need a proper multi-GPU server rather than a desktop install.

This local setup is separate from Ollama's official cloud tag, and remember the MiniMax Community License: self-hosting M3 for personal use or research is fine without permission, but a commercial deployment needs written authorization from MiniMax first.

Use MiniMax M3 in Your Own Scripts and Agents (API Access)

You're not limited to the interactive `ollama run` session. You can reach MiniMax M3's cloud tag through Ollama's REST API too, and any tool that already talks to a local Ollama instance, or to the OpenAI API format, can use it with a one-line model name change.

Generate an API Key

Visit ollama.com/settings/keys while signed in, click "Create API key", and copy the value. Set it as an environment variable:

export OLLAMA_API_KEY=your_api_key_here

Call MiniMax M3 from the Local Endpoint

curl http://localhost:11434/api/chat -d '{
  "model": "minimax-m3:cloud",
  "messages": [
    { "role": "user", "content": "Summarize the difference between MiniMax M2 and M3 in two sentences." }
  ],
  "stream": false
}'

Expected output (truncated):

{
  "model": "minimax-m3:cloud",
  "message": {
    "role": "assistant",
    "content": "MiniMax M2 is a smaller 230B-parameter MoE model that runs locally on Ollama with a 128K context window, while M3 scales up to 428B parameters with a new MSA attention mechanism, native multimodality, and up to 1M tokens of context, available only as a cloud tag."
  },
  "done": true
}

Call MiniMax M3 Directly from ollama.com

For a server or serverless function without a local Ollama install, send requests straight to ollama.com using your API key:

curl https://ollama.com/api/chat \
  -H "Authorization: Bearer $OLLAMA_API_KEY" \
  -d '{
    "model": "minimax-m3:cloud",
    "messages": [{ "role": "user", "content": "Hello" }]
  }'

Python Example

from ollama import Client

client = Client(host="https://ollama.com", headers={"Authorization": "Bearer " + api_key})

response = client.chat(
    model="minimax-m3:cloud",
    messages=[{"role": "user", "content": "Plan a migration from a REST API to gRPC for an internal microservice."}],
)
print(response["message"]["content"])

OpenAI-Compatible Endpoint for Existing Agent Tools

Ollama exposes an OpenAI-compatible layer at `http://localhost:11434/v1`, the same endpoint used in the Hermes Agent and OpenClaw setups. Point that configuration at `minimax-m3:cloud` instead of a local model name, and the agent runs on MiniMax M3's coding benchmarks and multimodal input without any other config changes:

model:
  default: minimax-m3:cloud
  provider: custom
  base_url: http://localhost:11434/v1
  context_length: 512000

Troubleshooting

Frequently Asked Questions