MiniMax M2.7 on Mac: Cloud API, Token Plan, and Local Limits

The three Mac access paths

MiniMax M2.7 targets coding, agent, and office workflows. On a Mac, three different access patterns need to be kept separate:

1. MiniMax API: Inference runs in MiniMax’s cloud and is billed either by token usage or through a Subscription Key.
2. Ollama Cloud: ollama run minimax-m2.7:cloud uses Ollama as a local client while the computation runs in the cloud.
3. Self-hosted weights: Hugging Face lists M2.7 as an open-weight model and points to SGLang, vLLM, and Transformers. That does not amount to an officially tested Apple Silicon or MLX workflow.

The speed and benchmark figures below come from MiniMax; no independent tokens-per-second measurements on Apple Silicon are available.

What makes MiniMax M2.7 different

The official MiniMax pages and model card position M2.7 as an agentic model for complex productivity tasks. They mention Agent Teams, complex Skills, dynamic tool search, software engineering, and office documents.MiniMax model page M2.7 announcement

Context and speed

Variant	Context window	Vendor-listed speed
MiniMax-M2.7	204,800 tokens total	approximately 60 output tokens/s
MiniMax-M2.7-highspeed	204,800 tokens total	approximately 100 output tokens/s

The API documentation explicitly says that the maximum token count refers to input and output combined. The speed figures are not measurements by this publication.API overview

How to read the benchmarks

MiniMax publishes the following results for M2.7, among others:

• SWE-Pro: 56.22%
• VIBE-Pro: 55.6%
• Terminal Bench 2: 57.0%
• SWE Multilingual: 76.5
• Multi SWE Bench: 52.7
• GDPval-AA: 1495 ELO
• MLE Bench Lite: 66.6% medal rate across 22 competitions

These are published vendor results. They show how MiniMax positions the model; they do not replace an independent test with your repository, agent harness, or Mac.MiniMax announcement Hugging Face model card

MiniMax also describes internal experiments in which M2.7 optimized a programming scaffold and worked with agent harnesses. Treat those statements as the manufacturer’s report, not as reproducible Mac experience.M2.7 announcement

Using MiniMax M2.7 on a Mac

1. API access

The official model page shows OpenAI-compatible HTTP access. The minimal example intentionally uses an environment-variable placeholder:

import os
import requests

response = requests.post(
    "https://api.minimax.io/v1/text/chatcompletion_v2",
    headers={"Authorization": f"Bearer {os.environ['MINIMAX_API_KEY']}"},
    json={
        "model": "MiniMax-M2.7",
        "messages": [{"role": "user", "content": "Hello"}],
    },
    timeout=60,
)
print(response.json())

The API overview also lists HTTP, the Anthropic SDK, and the OpenAI SDK. A pay-as-you-go API key and a Token Plan Subscription Key are different access and billing types.API overview

2. Ollama Cloud

Ollama lists M2.7 as minimax-m2.7:cloud:

ollama run minimax-m2.7:cloud

The :cloud suffix is the important part. Ollama is the local invocation or client; the model is not running as local Apple Silicon inference on the Mac. The Ollama page explicitly describes the access as commercially licensed cloud usage.Ollama Cloud listing

3. Self-hosted server

The Hugging Face model card provides weights and points to:

• SGLang
• vLLM
• Transformers
• ModelScope and an NVIDIA NIM endpoint as additional routes

The checked sources do not provide an official, tested MLX, llama.cpp, or Apple Silicon installation path for M2.7. The defensible Mac statement is therefore: cloud use is documented; local Mac execution remains a separate experiment whose hardware and runtime need validation. No specific RAM floor or Mac token rate is invented here.Hugging Face model card

Current API pricing and limits

Pay-as-you-go

MiniMax’s current pricing page separates standard API billing from monthly subscriptions. Prices below are per 1 million tokens:

Model	Input	Output	Cache read	Cache write
MiniMax-M2.7	$0.30	$1.20	$0.06	$0.375
MiniMax-M2.7-highspeed	$0.60	$2.40	$0.06	$0.375

A pay-as-you-go API key charges the Open Platform balance according to actual usage. That is not the same billing system as the Token Plan.Pay-as-you-go pricing API overview

For both M2.7 variants, MiniMax currently lists 500 requests per minute (RPM) and 20 million tokens per minute (TPM). These are API quotas, not a promise about latency or actual model speed.Rate limits

Token Plan

The Token Plan is a separate monthly subscription. Text, image, and speech resources share one usage quota according to the documentation, and M2.7 is included in the model coverage:

Plan	Monthly price	MiniMax’s described use
Plus	$20	personal projects and prototyping
Max	$50	daily coding with agents and multimodal work
Ultra	$120	heavy agent workflows and extended sessions

Quotas are shown in rolling five-hour and weekly windows. The documentation also describes approximately 3–4, 4–5, and 6–7 concurrent agents as orientation for the three tiers. This is product positioning, not an individual availability guarantee.Token Plan documentation

MiniMax also offers Credits packages. The checked documentation lists 1,000 credits for one US dollar, a 365-day validity period, and usage at the pay-as-you-go list price. If you only call M2.7 occasionally through the API, compare pay-as-you-go and Token Plan separately.Token Plan documentation

M2.7 or M3?

MiniMax now lists M3 as its newest M-series language model with a context window of up to 1,000,000 tokens. M2.7 remains available as a more specialized and lower-priced option in the API catalog; the API overview lists both model families explicitly.API overview

For a new workflow, the task matters:

• M2.7: worth considering when you need its documented coding and agent capabilities at the lower M2.7 price level.
• M2.7-highspeed: worth considering when the higher vendor-listed speed justifies the premium.
• M3: worth checking when very long context, multimodality, or the current M-series flagship matters more than M2.7 pricing.

OpenRoom and Agent Teams

MiniMax released OpenRoom, a Web GUI demo that shows agent interaction in an interactive environment. It is a product or research demo, not evidence of local inference on Apple Silicon. The graphic and video embedded below illustrate the Agent Teams approach described by MiniMax.M2.7 announcement

*Visualization of Agent Teams in M2.7; the graphic is a vendor illustration, not an independent Mac test.*

Privacy and the practical decision

With API and Ollama Cloud usage, code, prompts, and potentially files leave the Mac. Before using M2.7 in a production workflow, check at least:

• Which data your agent sends to the API or server-side tools.
• Whether the API key is supplied only through an environment variable or secret manager.
• Whether the actual billing path is pay-as-you-go, Subscription Key, or Credits.
• Whether a local model with a documented Apple Silicon path is better suited to private data.

For M2.7, the defensible short decision is: cloud first; local Mac use only as a specialist project that still needs validation.

Conclusion: Usable from the Mac — as a cloud model

MiniMax M2.7 works well on a Mac when “on Mac” means the local client, editor, or API workflow. In the standard setup, the model itself runs in MiniMax or Ollama Cloud. The official model card provides weights and server paths, but no tested Apple Silicon route. Anyone comparing costs must also separate pay-as-you-go from the Token Plan: API prices, Subscription Keys, quotas, and rate limits belong to different billing layers.

Sources and review date

Current as of August 25, 2026. Prices, limits, model catalogs, and product terms can change. Before paid use, re-check the linked original pages and your own MiniMax account.

• MiniMax M2.7 model page
• MiniMax M2.7 announcement
• MiniMax pay-as-you-go pricing
• MiniMax API overview
• MiniMax rate limits
• MiniMax Token Plan
• MiniMax M2.7 on Hugging Face
• Ollama MiniMax M2.7 Cloud