How to Self-Host MiniMax M2 with vLLM on One GPU

MiniMax M2 is open-weight, so you can run it yourself and pay for nothing but GPU time. The break-even against the hosted API comes down to one number: how many tokens per hour you actually push. Below about 5 million tokens a day, the API is cheaper. Above that, self-hosting starts to win.

Can MiniMax M2 run on one GPU?

Yes, in quantized form. The full weights need roughly 80GB of VRAM in 16-bit, which fits an A100 80GB or H100 80GB exactly. An 8-bit or 4-bit quantization drops that to about 40GB or 24GB, letting it run on a single RTX 4090 with 24GB — slower, but viable for development and low-traffic serving.

How do you serve it with vLLM?

The setup is short:

1. Install vLLM in a CUDA-matched environment: pip install vllm.
2. Launch the server: python -m vllm.entrypoints.openai.api_server --model minimax-m2 --tensor-parallel-size 1 --max-model-len 32768
3. Point any OpenAI SDK at http://localhost:8000/v1.
4. Set --gpu-memory-utilization 0.9 to leave headroom for the KV cache.

Because the endpoint is OpenAI-compatible, existing agent harnesses connect without code changes — the same benefit you get from the hosted route.

What throughput should you expect?

On an H100 in 16-bit, a well-tuned vLLM deployment serves several hundred output tokens per second per request and thousands pooled under batching. On a single 24GB card with 4-bit quantization, expect roughly 30-60 tokens per second per stream — fine for one developer, tight for a team.

Does self-hosting beat the API on price?

A rented H100 runs about $2-$3 per hour. If that hour pushes 5-10 million tokens, the effective rate beats the hosted API. If it sits idle most of the day, it does not. Teams serving one agent occasionally should stay on the API. Teams running continuous agent fleets, or with data that cannot leave the building, should self-host.

What are the real downsides?

You own the ops now. Model updates mean re-downloading tens of gigabytes. Context beyond 32k gets expensive in KV cache memory and may force a second GPU. And 4-bit quantization costs measurable reasoning quality, which matters on the hard planning steps that M2 is already weakest at. A pragmatic split: self-host the workhorse turns and call a stronger hosted model, such as what backs droid, only for planning.

Quick Summary

• MiniMax M2 needs ~80GB VRAM in 16-bit, ~24GB in 4-bit.
• Serve with vLLM’s OpenAI-compatible endpoint on port 8000.
• H100 serves hundreds of tokens/sec; a 4090 does 30-60 tokens/sec.
• Self-hosting wins above roughly 5-10M tokens per rented hour.
• You own updates, ops, and quantization quality loss.

Last updated September 2026.