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.

What is MiniMax M2?

M2 is MiniMax’s agentic model, tuned specifically for function calling and long tool chains rather than chat. It handles large context, so you can feed an entire repo slice plus prior tool results in one call. MiniMax positions it against mid-tier models like Claude Haiku at a fraction of the price, and the weights are open for self-hosting.

How do you call MiniMax M2?

Two practical routes:

• Hosted API: create a MiniMax account, take an API key, and call the chat completions endpoint with your model name.
• Self-host: pull the open weights, serve with vLLM on a single 80GB GPU, and point your client at the local endpoint.

The API is OpenAI-compatible, so most SDKs work by changing the base URL and model string. No new client library is needed.

How do you use it in an agent loop?

1. Define your tools in the standard JSON schema format.
2. Send the task plus tool definitions in one call.
3. Execute the returned tool calls, append results, and loop.
4. Cap turns at 15-25 to bound cost and avoid runaway loops.

Because M2 was trained on tool chains, it rarely invents argument names the way general chat models do. That reliability is the main reason teams swap it into existing agent harnesses.

How much does MiniMax M2 cost?

Hosted pricing sits in the low single digits per million tokens, far under a frontier model. A 20-turn coding agent session averaging 40k tokens per turn costs a few cents. Self-hosting trades that for GPU time — a rented 80GB card runs $1.50-$2.50 per hour, which only pays off above steady high volume.

Where does MiniMax M2 fall short?

M2 is not a top-tier reasoner. On subtle architecture calls or tricky concurrency bugs it produces shallower plans than Claude Sonnet or GPT-class models. It is also weaker at long-form prose. Use it as the workhorse in a loop, and escalate the hardest planning step to a stronger model — the same hybrid pattern that works with opencode.

Quick Summary

• MiniMax M2 is an agent-tuned, open-weight model with a long context window.
• Access via the OpenAI-compatible MiniMax API or self-host with vLLM.
• Hosted cost is low single digits per million tokens — cents per agent session.
• Strong at tool calling, weaker at deep reasoning and long prose.
• Best as a cheap workhorse with escalation for hard planning.

Last updated September 2026.