MiniMax M3.1 Flash on Ollama: Cloud Model, Weights Status and Local Limits

Run it in one command

Ollama lists the model as a cloud-hosted entry:

ollama run minimax-m3.1-flash:cloud

The same model over the HTTP API:

curl https://ollama.com/api/chat \
  -H "Authorization: Bearer $OLLAMA_API_KEY" \
  -d '{
    "model": "minimax-m3.1-flash",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Python and JavaScript clients exist too, all pointing at https://ollama.com with the same bearer-token pattern.

What the Ollama entry says

Field	Value
Parameters	428B
Context window	976K (cloud variant)
Modalities	vision, tools, thinking
Price	$0.15 in / $0.015 cached / $0.60 out per 1M tokens
Free window	Shown as free until 22 October 2026

The prices match MiniMax’s own list price, so using Ollama as the front end does not change the token economics — only the billing path and the free window.

Launching coding agents from Ollama

Ollama can start an agent already wired to this model:

ollama launch claude --model minimax-m3.1-flash:cloud
ollama launch opencode --model minimax-m3.1-flash:cloud
ollama launch hermes --model minimax-m3.1-flash:cloud
ollama launch openclaw --model minimax-m3.1-flash:cloud

The OpenCode and agents page compares this route with talking to MiniMax directly.

Are the weights open yet?

Short answer: no, not as of 21 October 2026. Four sources agree, and one marketing phrase does not:

Source	What it says
MiniMax launch post	“Model weights will be released later”
MiniMax homepage copy	Calls it “an ultra-efficient open-weight model” — marketing language, not a release
Hugging Face	`MiniMaxAI/MiniMax-M3.1-Flash` returns 404 — the repository does not exist
Artificial Analysis	`isOpenWeights = false`, `openSourceCategorization = proprietary`, no weights URL
MiniMax self-hosting docs	The local-deployment guides cover M3, M2.7, Music 3 and H3 — not M3.1

So the honest position is: the weights have not been published, and the documentation that would tell you how to self-host M3.1 does not exist yet.

If you need it locally today

Option	Trade-off
Ollama cloud (`:cloud`)	One command, no hardware, same list price
MiniMax API	Direct, full parameter control, see API guide
Aggregators (OpenRouter, Fireworks AI, Vercel AI Gateway)	Convenience and failover, small routing premium
Wait for the weights	Only path to local inference, at 428B parameters of hardware

For scale: a 428B-parameter model in a mixture-of-experts configuration is still a multi-GPU machine, and MiniMax’s own M3 self-hosting guide uses SGLang. We are not going to promise a local recipe for a release that has not happened — but if you want to watch for it, the Hugging Face organisation page and the release notes are the two places it would appear.

What we could not verify

We have not run the model locally, because there is no local weight to run. Everything on this page about the cloud variant comes from Ollama’s own listing and from the model documentation.

Next: OpenCode and agents for the agent workflows, or pricing for what these calls actually cost.

MiniMax M3.1 Flash on Ollama and the Open-Weights Question