How to Run MiniMax M3 on Canopy Wave's Shared Endpoint

This post covers what makes MiniMax M3 special from a technical perspective and shows you exactly how to start using it on our platform.

MiniMax M3 is a 427-billion-parameter Mixture-of-Experts model that pushes the boundary of what open models can do. It is offered in BF16 quantization and combines extreme MoE sparsity with MiniMax's proprietary MiniMax Sparse Attention (MSA) architecture — a high-performance sparse attention mechanism designed for million-token contexts. MSA delivers roughly 9× prefill and 15× decode speedups compared to the previous generation at 1M context, reducing per-token compute to about 1/20 while preserving model quality.

M3 reaches industry-leading levels in coding and agent evaluations. It features autonomous task decomposition, tool calling, and multi-step reasoning. The code it produces aims to be directly deliverable, not just "it runs but still needs human fixes."

Powered by the proprietary MiniMax Sparse Attention (MSA) architecture, the API supports up to a 1M-token context window, with a guaranteed minimum of 512K tokens available. The 1M context serves as the foundational infrastructure for long-horizon Agents, long-horizon Coding, and long-video understanding.

M3 is a native multimodal model. It undergoes mixed-modality training from step zero, achieving deep alignment between textual and visual semantic spaces. Multimodality is a core capability baked into the model's foundation, not a shallow layer added later.

On the BrowseComp agent benchmark, M3 scores 83.5, surpassing Opus 4.7 (79.3) and demonstrating strong autonomous browsing and information retrieval capabilities.

Previously, only a handful of closed-source models could simultaneously deliver frontier-level coding, a million-token context, and native multimodality. M3 is the first model to bring this complete set of frontier capabilities into the open world.

It comes with:

• A native 1-million-token context window

• Strong multimodal capabilities (text, images, and video)

• Excellent performance on long-horizon coding, agentic workflows, knowledge work, and complex reasoning tasks

Its efficient architecture and support for modern precision formats (including BF16) make high-quality inference far more accessible than a traditional dense checkpoint of similar scale.

Running a model of MiniMax M3's scale yourself requires significant GPU resources and operational effort. Canopy Wave's Shared Endpoint removes that complexity while giving you production-ready access:

• One-click migration with zero code changes — Fully OpenAI-compatible. Simply replace the base_url and API key to switch. Existing code, tools, and SDKs require almost no modification.

• Zero infrastructure management — No need to provision GPUs, handle scaling, or worry about cold starts. Just call the API and start building.

• Low latency, high throughput — Optimized inference stack designed for concurrent workloads, delivering fast and stable responses.

• SOC 2 certified — Passed rigorous international security audits with strict controls over data access and storage processes.

• Privacy-first architecture — Built with privacy at the core and enforced zero-data-retention policies, ensuring your data remains fully under your control.

• Production-grade high availability — Engineered for production workloads with high-availability guarantees and continuous uptime.

• Elastic scaling path — Start on the Shared Endpoint and seamlessly migrate to a Dedicated Endpoint as your business grows, with capacity fully assured.

This combination makes Canopy Wave one of the simplest and most reliable ways to use MiniMax M3 in real applications.

Canopy Wave's Shared Endpoint is fully OpenAI-compatible. You can use the exact same tools and libraries you already know.

• Set your key as an environment variable:

export CANOPYWAVE_API_KEY="your_api_key_here"

• curl (for the first example)

• Python 3.10 or higher, and the openai package (for the Python examples)

curl -X POST https://inference.canopywave.io/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $CANOPYWAVE_API_KEY" \
-d '{
 "model": "minimax/minimax-m3",
 "messages": [
   {"role": "user", "content": "tell me a story"}
 ],
 "max_tokens": 1000,
 "temperature": 0.7
}'

from openai import OpenAI
import os

BASE_URL = "https://inference.canopywave.io/v1"
API_KEY = os.environ.get("CANOPYWAVE_API_KEY")

client = OpenAI(api_key=API_KEY, base_url=BASE_URL)

response = client.chat.completions.create(
    model="minimax/minimax-m3",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "please tell me a story."}
    ],
)

print(response.choices[0].message.content)

• Use a clear system prompt when you need specific behavior (coding style, role-playing, agent instructions, etc.).

• Take advantage of the full 1M context window for large codebases, long documents, multi-turn agent workflows, or long video understanding.

• The model supports thinking/reasoning modes (enabled, adaptive, or disabled) via the appropriate parameter where available — experiment to balance depth and latency.

• The model ID must be exactly minimax/minimax-m3.

MiniMax M3 is one of the most capable open models available today for advanced coding, long-context agentic work, multimodal understanding, and complex reasoning. It is the first to bring the complete set of frontier capabilities — coding excellence, million-token context, and native multimodality — into the open world. With Canopy Wave's Shared Endpoint, you can start using it immediately — no cluster to manage, no cold starts, and full OpenAI compatibility.