MiniMax M3 Now Available on TheRouter.ai

MiniMax released M3 on June 1, 2026. It's now available on TheRouter.ai as minimax/minimax-m3.

The headline numbers: 1M token context window, native image and video input, 59.0% on SWE-Bench Pro (self-reported), and launch pricing of $0.30 input / $1.20 output per million tokens. Open weights are announced, targeting Hugging Face and GitHub.

What Changed from M2.7

MiniMax M2.7 was a capable text model with a long context window. M3 is a different class of model — the architecture was redesigned to reduce compute to roughly 1/20th of M2.7 while expanding capability across three dimensions:

1. Multimodal input: M3 natively understands images and video frames in the same request as text. M2.7 was text-only.
2. Coding performance: M3 scores 59.0% on SWE-Bench Pro, placing it in the same tier as frontier coding models. M2.7 had no comparable benchmark.
3. Context guarantee: M3 has a 1M token nominal context window with a 512K minimum guaranteed by SLA. M2.7's context was longer in nominal terms but without an SLA floor.

The compute reduction is significant for deployment: lower inference cost translates directly to lower pricing, which is why the launch price for M3 is aggressive even by aggregator-market standards.

Key Specifications

Pricing

The launch price makes M3 one of the most cost-efficient multimodal models available. At $0.30 input, you can process roughly 3.3 billion tokens per dollar — comparable to what fast text-only models cost on most platforms.

Access M3 on TheRouter.ai with a single API key alongside every other model in the catalog. No separate MiniMax account required.

Quick Start

from openai import OpenAI

client = OpenAI(
    api_key="your-therouter-api-key",
    base_url="https://api.therouter.ai/v1"
)

# Text-only request
response = client.chat.completions.create(
    model="minimax/minimax-m3",
    messages=[
        {"role": "user", "content": "Explain the trade-offs between RAG and fine-tuning for a production QA system."}
    ],
    max_tokens=1024
)
print(response.choices[0].message.content)

Image input

response = client.chat.completions.create(
    model="minimax/minimax-m3",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {"url": "https://example.com/architecture-diagram.png"}
                },
                {
                    "type": "text",
                    "text": "Identify any single points of failure in this architecture diagram."
                }
            ]
        }
    ],
    max_tokens=2048
)
print(response.choices[0].message.content)

Long-context document analysis

M3's 1M token context fits roughly 750,000 words — a large codebase, a full legal contract corpus, or an entire product changelog history in a single prompt. Here's how to use it for a long document analysis task:

with open("full_codebase.txt", "r") as f:
    codebase_content = f.read()

response = client.chat.completions.create(
    model="minimax/minimax-m3",
    messages=[
        {
            "role": "system",
            "content": "You are a senior engineer reviewing a codebase for security vulnerabilities and architectural problems."
        },
        {
            "role": "user",
            "content": f"Review this entire codebase and produce a prioritized findings report:\n\n{codebase_content}"
        }
    ],
    max_tokens=4096
)
print(response.choices[0].message.content)

What 1M Context Actually Unlocks

The 512K SLA guarantee means you can build applications that reliably load large contexts without worrying about provider-side truncation during peak load. Practical use cases:

• Full codebase review: A medium-sized TypeScript project (~100K lines) fits well within 512K tokens, giving M3 complete visibility across every file in one pass.
• Legal and compliance document analysis: Load an entire contract corpus — NDAs, MSAs, SOWs — and query relationships across documents without chunking.
• Long conversation memory: Store full conversation histories for customer support agents or personal assistants without summary compression.
• Multi-document research synthesis: Send 50+ academic papers or reports in a single prompt and ask for cross-document synthesis.

Native Multimodal Input

M3 accepts image URLs (HTTP/HTTPS) and base64-encoded images in the same format as other OpenAI-compatible multimodal models. Video is accepted via the standard image_url type — pass individual video frames or a video URL depending on your provider configuration.

This makes M3 a drop-in replacement for vision workloads currently running on models like GPT-4o or Claude 3.5 Sonnet, at a fraction of the cost during the promotional period.

Routing on TheRouter.ai

TheRouter routes M3 via both MiniMax's direct API and SiliconFlow, giving you automatic failover without changing your application code. If one upstream is degraded, traffic shifts to the healthy route transparently.

You can also pin to a specific upstream if needed using TheRouter's routing headers — but for most production workloads, the default automatic routing is the right choice.

Getting Started

1. Sign up at therouter.ai and get an API key.
2. Set base_url="https://api.therouter.ai/v1" in your client.
3. Use model="minimax/minimax-m3" in your request.

No additional configuration required. The same key accesses every model in the TheRouter catalog.