MiniMax M3 vs M2.7: Benchmarks, Specs & Migration

MiniMax now lists only two current text models: MiniMax-M3 and MiniMax-M3.1-Flash. M2.7 and M2.5 are on their way out, with retirement set for November 18, 2026. If your product runs on M2.7, you need to decide what to move to before that date. This guide compares M3 and M2.7 using only MiniMax's published numbers, explains what actually changes in your requests, and walks through the switch. On APIMart, both models are available through the MiniMax API with the same OpenAI-compatible request format, so the move is mostly a configuration change.

Key Takeaways

• MiniMax M3 is a natively multimodal model that accepts text, images and video, with a 1,000,000-token context window and up to 524,288 output tokens per response. M2.7 tops out at a 204,800-token context.
• On MiniMax's own benchmark chart, M3 only edges M2.7 on SWE-Bench Verified (80.5 vs 79.9) but pulls well ahead on agent work: Terminal Bench 2.1 66.0 vs 51.1, MCP Atlas 74.2 vs 49.4.
• MiniMax stopped offering M2.7 and M2.5 to new users on October 19, 2026 and retires them on November 18, 2026. After that date, pay-as-you-go requests to them stop working.
• Migrating is mostly a one-line change to minimax-m3, but thinking and sampling behave differently: M3 lets you turn thinking off, M2.x does not, and M3 always runs at temperature 1.
• As of November 4, 2026, APIMart Gold members pay the same $0.24 input / $0.96 output per 1M tokens for both models.

MiniMax M3 vs M2.7 at a Glance

Sources: MiniMax M3 model card, MiniMax M2.7 config, MiniMax model list and MiniMax Chat Completions reference.

What Changed from M2.7 to M3

Native multimodality

MiniMax describes M3 as a "frontier multimodal coding model" that was trained on mixed modalities from the first step. In practice, that means one model can read a screenshot, a scanned PDF page or a short video alongside your prompt. MiniMax reports 78.1 on MMMU-Pro, 91.6 on OmniDocBench and 84.6 on Video-MMMU for M3. The chart has no M2.7 scores for these tests, so there is no like-for-like multimodal comparison to make.

A context window five times larger

M3 handles up to 1,000,000 tokens of context, compared with 204,800 for M2.7. That is the difference between feeding a model one large repository module and feeding it most of a codebase, or between a long contract and a full data room. A bigger window does not guarantee better recall at the far end, so test retrieval on your own long inputs before you rely on it.

Longer outputs and controllable thinking

M3 can return up to 524,288 tokens in one response, and MiniMax recommends 131,072 for most requests. M2.x models cap at 204,800, with 65,536 recommended. Thinking changed too. M3 uses adaptive thinking by default, deciding on its own when extra reasoning helps, and you can switch it off for faster replies. On M2.x, thinking is always on, and a disabled setting is accepted but ignored.

Benchmarks: MiniMax M3 vs M2.7

All numbers below come from the benchmark chart on the MiniMax M3 model card. They are MiniMax's own results, mostly run on internal infrastructure with Claude Code as the scaffolding, not independent third-party tests.

Source: MiniMax M3 model card (MiniMax-reported).

Two patterns stand out. On classic single-repository bug fixing, the gap is small: 0.6 points on SWE-Bench Verified and 2.8 on SWE-Bench Pro. If your workload is mostly "fix this issue in this repo", M2.7 was already close to M3. The gap opens up on multi-step, tool-heavy work. M3 leads by 24.8 points on both MCP Atlas and Claw-Eval, by 14.9 on Terminal Bench 2.1, and by 26.6 on SWE Atlas-QnA. Those are the benchmarks closest to agents that call tools, run commands and answer questions about a whole codebase.

Pricing

As of November 4, 2026, APIMart Gold members pay $0.24 per 1M input tokens and $0.96 per 1M output tokens for minimax-m3, minimax-m2.7 and minimax-m2.5 alike. Switching to M3 does not change the per-token rate. Your bill can still move, because M3's adaptive thinking and longer outputs change how many tokens a request uses. Track tokens per request for a few days after the switch. Current rates for every MiniMax model are on the MiniMax pricing page.

Availability and Retirement Timeline

• October 19, 2026: MiniMax stopped offering M2.5 and M2.7 to new users.
• November 18, 2026: M2.5 and M2.7 are retired. MiniMax says pay-as-you-go API requests to them stop working after this date.
• Already retired: MiniMax-M2.1, MiniMax-M2 and M2-her.

As of November 4, 2026, APIMart still lists minimax-m2.7 and minimax-m2.5. Because MiniMax is retiring the underlying models, plan to finish your switch before November 18.

How to Migrate from M2.7 to M3

The request format stays the same, so most teams can migrate in an afternoon. Work through these five changes:

1. Change the model ID from minimax-m2.7 (or minimax-m2.5) to minimax-m3.
2. Decide on thinking. M2.7 always reasoned before answering. M3 defaults to adaptive thinking. If you want fast, direct answers for chat or classification, send "thinking": {"type": "disabled"}. With the OpenAI Python SDK, put it inside extra_body.
3. Drop temperature tuning. M3 always runs at temperature 1, and any other value has no effect. If you relied on a low temperature for stable output, use clearer instructions or structured output instead.
4. Revisit max_completion_tokens. M3 accepts up to 524,288. Start with MiniMax's recommended 131,072 for long tasks, and keep a lower cap for chat to control cost.
5. Re-run your evals. Compare quality, latency and tokens per request on real traffic before you move 100% of it.

A minimal request with thinking turned off looks like this:

curl https://api.apimart.ai/v1/chat/completions \
  -H "Authorization: Bearer $APIMART_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-m3",
    "messages": [{"role": "user", "content": "Summarize this pull request in three bullet points."}],
    "thinking": {"type": "disabled"},
    "max_completion_tokens": 4096,
    "stream": false
  }'

The same call with the OpenAI Python SDK:

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.apimart.ai/v1", api_key=os.environ["APIMART_API_KEY"])

response = client.chat.completions.create(
    model="minimax-m3",  # previously "minimax-m2.7"
    messages=[{"role": "user", "content": "Summarize this pull request in three bullet points."}],
    max_completion_tokens=4096,
    stream=False,
    extra_body={"thinking": {"type": "disabled"}},  # remove this line to keep adaptive thinking
)
print(response.choices[0].message.content)

For the full parameter list, see the APIMart Chat Completions reference.

How MiniMax M3 Compares with Kimi, GLM and DeepSeek

The same MiniMax chart includes other open-weight models. The figures below are copied from that chart; a dash means MiniMax did not report a score. MiniMax notes that the NL2Repo scores for DeepSeek, Kimi and GLM were taken from Qwen's published results, while most other columns were tested by MiniMax through each vendor's official API.

Source: MiniMax M3 model card (MiniMax-reported).

On MiniMax's numbers, M3, DeepSeek V4 Pro and Kimi K2.6 Thinking are within 0.4 points of each other on both SWE-Bench Verified and BrowseComp, while GLM 5.1 Thinking has no SWE-Bench Verified score and trails on BrowseComp at 79.3. Kimi K2.6 Thinking is also slightly ahead on NL2Repo. M3's clearest leads are on Terminal Bench 2.1 and Claw-Eval. Because these are vendor-reported results, treat them as a shortlist, not a verdict, and run your own tasks before you pick one.

Which One Should You Use?

• Starting a new project: use M3. MiniMax no longer offers M2.7 to new users, and it retires on November 18.
• Running M2.7 in production: move to M3 now, while you still have time to compare outputs side by side.
• Need image or video input: M3 is the MiniMax text model with confirmed multimodal input.
• Mostly single-repo bug fixing: expect a modest quality change; the bigger gains are in tool-heavy agent work.

FAQs

What is the difference between MiniMax M3 and M2.7?

M3 is MiniMax's current flagship. It accepts text, images and video, supports a 1,000,000-token context and up to 524,288 output tokens, and lets you turn thinking off. M2.7 supports a 204,800-token context, always thinks before answering, and is being retired.

Is MiniMax M2.7 being retired?

Yes. MiniMax stopped offering M2.7 and M2.5 to new users on October 19, 2026 and retires both on November 18, 2026. After that, pay-as-you-go requests to them stop working.

How do I migrate from MiniMax M2.7 to M3?

Change the model ID to minimax-m3, decide whether to keep adaptive thinking or send "thinking": {"type": "disabled"}, remove temperature tuning because M3 always uses 1, and review max_completion_tokens. Then re-run your evaluation set before switching all traffic.

How does MiniMax M3 perform on benchmarks?

On MiniMax's own chart, M3 scores 80.5 on SWE-Bench Verified, 59.0 on SWE-Bench Pro, 66.0 on Terminal Bench 2.1 and 83.5 on BrowseComp. These are vendor-reported results, not independent tests.

Is MiniMax M3 open source?

MiniMax publishes M3's weights on Hugging Face under the MiniMax Community License. It is an open-weight model with its own license terms, so check the license before commercial self-hosting.

How much does MiniMax M3 cost?

As of November 4, 2026, APIMart Gold members pay $0.24 per 1M input tokens and $0.96 per 1M output tokens for MiniMax M3, the same rate as M2.7. Check the MiniMax model page on APIMart for current prices.

What to Watch Next

MiniMax has already released MiniMax-M3.1-Flash, which it describes as a multimodal coding model with a 1M context window and tunable thinking depth. APIMart does not offer it yet, and we will not guess at a launch date. In the meantime, the practical step is clear: move M2.7 workloads to M3 before November 18, and measure quality, latency and tokens per request as you go.