MiniMax M2.7 API setup: base URL, exact id, first request.

Quick answer: two values, one capitalization trap.

Every OpenAI-compatible SDK takes a base URL and a key. Set the base URL to https://api.apisrouter.com/v1, set the model field to MiniMax-M2.7, and the request is the same /v1/chat/completions payload as every other chat model in the catalog. The id is the one gotcha: it is capitalized, MiniMax-M2.7, unlike most catalog ids, which are lowercase. Model ids are exact strings, so minimax-m2.7 returns model-not-found rather than routing anywhere. Copy the spelling from the /v1/models listing rather than from habit, and the wiring is done in one request.

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $APISROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-M2.7",
    "messages": [{"role": "user", "content": "Plan the steps to add pagination to a REST endpoint, then write the diff."}]
  }'

What MiniMax M2.7 actually is.

MiniMax released M2.7 on March 18, 2026, positioning it for agentic workflows, software engineering, and office tasks. Architecturally it is a sparse Mixture-of-Experts reported at about 230 billion total parameters with roughly 10 billion active per token, which is the design choice behind its economics: a small active footprint keeps serving costs low, and MiniMax's published rate reflects that. The launch's most-quoted claim is the self-evolution story: MiniMax described M2.7 as the first of its models to participate in its own development, with an internal version autonomously running more than a hundred optimization rounds over the reinforcement-learning pipeline that produced it. It is a striking claim and a vendor claim; treat it as positioning rather than a spec. The weights are public: coverage in April 2026 reported the open-weight release on Hugging Face under a Modified-MIT license, with one caveat that matters commercially and got its own news cycle: the license's commercial-use clause requires MiniMax's prior written authorization, so self-hosting it inside a paid product is a licensing conversation, not a download.

Quickstart: Python and JavaScript with the OpenAI SDK.

Both official OpenAI SDKs accept a custom base URL at client construction, and the model field passes through as a plain string. Nothing MiniMax-specific is required: chat completions, streaming, and the standard tools array work the way they do for every other id on the endpoint. Keep the key in an environment variable, and pin the id in one config constant, capitalization included, so the string lives in exactly one place.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APISROUTER_API_KEY"],
    base_url="https://api.apisrouter.com/v1",
)

resp = client.chat.completions.create(
    model="MiniMax-M2.7",
    messages=[
        {"role": "system", "content": "You are a careful software engineer."},
        {"role": "user", "content": "Review this function for edge cases and propose a fix: ..."},
    ],
    max_tokens=4000,
)
print(resp.choices[0].message.content)
print(resp.usage)  # prompt_tokens / completion_tokens per request

Agent and coding strengths, with the sources attached.

The practical read: M2.7's case is efficiency at agent work, competitive multi-step behavior from a 10B-active design at one of the lowest rates in the catalog's current generation. Loops that run hundreds of steps care about exactly that ratio, and the usage log prices it on your own workload within an afternoon.

• Agentic coding: MiniMax reported 56.2 on SWE-bench Pro and 57.0 on Terminal-Bench 2.0 at launch, with 76.5 on SWE-bench Multilingual, a benchmark family the M2 line has historically led among open models.
• Agentic browsing and tool use: a reported 77.8 on BrowseComp and launch materials emphasizing multi-step loops: inspect logs, reason about code, modify files, call tools, iterate.
• Office and artifact work: MiniMax's launch framing put document and office-task generation alongside coding, unusual among agent-focused releases and relevant if your loops produce reports rather than diffs.
• The caveat that applies to all of the above: these are vendor-reported launch figures. Independent leaderboards have M2.7 competitive in the open-weight agent class rather than atop it; the GLM-5.2 comparison linked below carries the strongest current rival's numbers.

Context: what the catalog serves vs what the vendor publishes.

One discrepancy to state plainly, because budgeting depends on it. The MiniMax-M2.7 catalog entry here lists a 1M-token context window. MiniMax's own model documentation describes M2.7 with a context on the order of 200K tokens (196,608 in the vendor's spec). Those numbers do not agree, and this page will not pretend they do. The conservative engineering choice is to budget prompts against the vendor's 200K figure until a long-context test on this route proves otherwise: send a deliberately oversized prompt once and observe where truncation or rejection actually lands. If your workload never approaches 200K, the discrepancy costs you nothing; if it does, five minutes of testing beats an assumption in either direction. Whichever ceiling applies, input tokens bill on every request, so the usual long-context discipline holds: scope prompts to the task, summarize long-lived state, and treat giant contexts as spend-worthy exceptions rather than defaults.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Verify the endpoint and debug the first request.

List the served models first and copy the exact id from the response; for this model the capitalization is the most common failure. The rest of the first-request list is standard. A 401 means the key was not exported into the shell running your code. A 404 on the path means the base URL is missing its /v1 suffix, since SDKs append /chat/completions to the base they are given. A response truncated with finish_reason=length on multi-step work means max_tokens is too tight for a model that reasons through plans; raise it with headroom before judging output quality. For what M2.7 costs against its published rate, and how the numbers work at volume, the dedicated MiniMax M2 pricing page linked below carries the worked examples; this page stays on setup so the two do not repeat each other.

curl -s https://api.apisrouter.com/v1/models \
  -H "Authorization: Bearer $APISROUTER_API_KEY" | grep -i minimax

FAQ