Reference

MiniMax M3.1 Flash FAQ and Troubleshooting

Nineteen questions with one-sentence answers, each with the source it came from.

Reviewed

Model basics

1. Can I still use MiniMax-M3.1-Flash-Preview?

Yes, technically: the preview identifier was retired when the model went generally available on 19 October 2026, and requests using the old name are still routed to the current model. But new code should use MiniMax-M3.1-Flash. Source: M Plan overview documentation.

2. Is the context window really 1 million tokens?

Yes — 1,000,000 tokens of input, with a maximum output of 524,288 tokens. These are different limits and pages often blur them. In Claude Code the effective window also depends on the harness setting CLAUDE_CODE_AUTO_COMPACT_WINDOW and the [1m] model suffix. Source: model invocation docs, Ollama listing.

3. When exactly was it released?

Generally available on 19 October 2026. A preview version shipped inside MiniMax Code on 27 September 2026 and has since been retired. Source: release notes, launch post.

4. What is the difference from MiniMax-M3?

M3.1 Flash adds always-on reasoning with an adjustable effort level, a broader agentic and cybersecurity capability set, and lower task cost. M3 remains available and, unlike M3.1 Flash, doubles in price above 512k input tokens. Source: GA page comparison table, pricing page.

5. What other MiniMax models exist, and which should I use?

Text is M3.1 Flash and M3. Video is the H3 family (H3, H3-Max, Hailuo-2.3, Hailuo-2.3-Fast, Hailuo-02). Speech is speech-2.8-hd/turbo with older 2.6 and 02 variants. Image is image-01 and image-01-live. The Music generation line is discontinued. For a coding agent, use MiniMax-M3.1-Flash. The full table is on the model list page. Source: MiniMax model list documentation.

6. Is MiniMax-M2.7 / M2.5 going away?

Yes. From 19 October 2026 they are closed to new users, and on 18 November 2026 they are fully retired for pay-as-you-go users. M Plan and Token Plan subscribers are automatically rerouted to M3.1 Flash from 19 October, and the old model names keep working. Source: MiniMax model list retirement notice.

Cost

7. Why is my bill higher than expected?

Three likely causes, in order of probability:

  1. Reasoning effort defaults to max. The API accepts low, medium, high, xhigh and max; omitting the parameter gives you max.
  2. Prompt cache misses. Cached input costs $0.015 per million instead of $0.15.
  3. Long context. M3.1 Flash does not step its price up with context, but every token still costs.

See pricing for the arithmetic. Source: model invocation docs, pricing page.

8. Is there a free tier?

The OpenRouter free launch period ended at 12:00 UTC on 22 October 2026 and the :free variant was retired; Ollama showed the model as free until the same date. After that, list price applies. Source: OpenRouter model page, Ollama listing.

9. What does a task actually cost?

Artificial Analysis measures a weighted cost per task of $0.22 and 615 seconds per task for M3.1 Flash on its benchmark battery, versus $0.571 for DeepSeek V4.1 Flash. Your own mix will differ. Source: Artificial Analysis model pages.

Capabilities

10. Does it accept images and video?

Yes for input — text, image and video — with text output. Note that video attachments can consume a large share of a 1M-token context. Source: model invocation docs, GA page.

11. Does it support function and tool calling?

Yes, and it is one of the model’s headline strengths: MiniMax documents tool use and interleaved thinking as first-class features, and BankerToolBench (76.2) and Automation-Bench are agentic tool-use evaluations. Source: tool use documentation, launch post.

12. Are the weights open?

Not yet, as of 21 October 2026. The launch post says weights will be released later; the Hugging Face repository returns 404; Artificial Analysis records isOpenWeights = false; and MiniMax’s own self-hosting guides do not cover M3.1. See the Ollama page for the full four-source check.

Setup

13. Which coding tools are supported?

MiniMax publishes guides for MiniMax Code, Claude Code, Codex, Cursor, OpenCode, Pi, Hermes Agent, OpenClaw, generic OpenAI-compatible tools, an MCP server with web search and image understanding, and the MiniMax CLI. We have detailed guides for Claude Code, Codex, Cursor and OpenCode.

14. Why does my agent still use the wrong model?

In Claude Code and Codex, the usual cause is an inherited environment variable. ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN take priority over the settings file, and exporting them in ~/.bashrc or ~/.zshrc makes the fix look like it did not work. Unset them and remove the lines.

15. What does “model not found” mean?

Almost always a stale identifier. Check for MiniMax-M3.1-Flash-Preview in your configuration and replace it with MiniMax-M3.1-Flash (or MiniMax-M3.1-Flash[1m] in Claude Code).

Subscription

16. I subscribed to M Plan but my account balance is still dropping.

Your tool is configured with a pay-as-you-go API Key. A Subscription Key only ever spends M Plan usage and then credit packs — it never touches your account balance. Swap it for the Subscription Key from Plan Details. Source: M Plan FAQ.

17. When does M Plan quota reset?

Not monthly. Text, image, audio and other non-video models are governed by a 5-hour window and a weekly window; video models have only the weekly window. Both windows start with your first use — the 5-hour window lasts 5 hours, the weekly one 7 days — and they reset independently. Unused usage does not roll over. Source: M Plan FAQ.

18. What is a credit pack, and is it cheaper than pay-as-you-go?

No. 1,000 Credits = $1, exactly the pay-as-you-go list price, and credit packs are valid for one year. They are consumed only after your plan usage runs out, starting with the pack that expires soonest. A credit pack is a different wrapper for the same money, not a discount. Source: M Plan FAQ.

19. Why does mmx auth login keep returning 401?

The CLI usually detects the service region from your API Key, so a persistent 401 means detection failed. Set it manually with mmx config set --key region --value global or --value cn, then check with mmx auth status. Source: M Plan FAQ.

Where to go next

Sources and version notes

Checked October 21, 2026. We label official claims, third-party measurements, our own testing and community reports separately — see the sources policy for what each label means.

Spotted something out of date? Suggest a correction →

Your next step

API referenceTwo compatible endpoints, one model ID, an effort dial that defaults to max, and a 1M-token context window.Pricing & costFlat per-million-token pricing with no long-context tier, plus the cache and effort settings that actually decide your bill.Claude CodeThe working settings.json, the environment variables that silently override it, and the model ID that unlocks the long context.CodexA provider block in config.toml, one extra catalog file, and a version floor that has bitten people.OllamaA one-command cloud model today, 428B parameters, and no local weights yet — with three independent sources on that last point.M PlanThree tiers, one text model in all of them, two usage windows, and a break-even rule you can compute yourself.Benchmarks & trustThe official numbers, the third-party numbers, the known flaws in one of the headline benchmarks, and what is simply not published.Model listOne table of every live model, the retirement timeline for the M2.x family, and which one to point your agent at.