dalao.ai

MiniMax

Multimodal foundation models and AI-native productsReference profile

About MiniMax

MiniMax, a foundation-model company it says was founded in early 2022, has the mission “Intelligence with Everyone.” It builds text (M3 / M3.1 Flash), video (H3 / Hailuo), speech (Speech 2.8) and music models, and offers MiniMax Code, Design, Audio, Talkie (Xingye in China) and a developer platform. Hosted products and APIs are closed commercial services; M3 and M3.1-Flash weights use the MiniMax Community License. H3-Base video weights use a separate MiniMax H3 Community License that lists the EU, UK, South Korea and the US as Excluded Territories; use there requires a separate license.

EditorialUpdated 13 min read

What problem it solves

MiniMax sets out to cover "text agents, video generation and speech synthesis from one company, with both consumer products and developer APIs". The official About page says the company was founded in early 2022; the Chinese investor-relations page says "since its founding in early 2022" and frames the mission as "co-creating intelligence with everyone". Operating figures such as user numbers and country coverage are vendor claims (the IR page says it has served more than 300M individual users from over 230 countries and regions, and more than 2M enterprise customers and developers from over 100 countries and regions).

Products and platform (2026-10-25). On the consumer and creator side there are MiniMax Agent / Code, Hailuo video, MiniMax Audio, Talkie and others; on the developer side there is the open platform, with separate China (platform.minimax.cn) and international (platform.minimax.io) sites billed in different currencies; M3 and M3.1-Flash weights are released on Hugging Face. The pricing overview splits capabilities into language models, video, speech, image and music, plus MCP / server tools (such as web_search); there are also voice packages and video packages (Hailuo), and the international site adds an Audio Subscription. The official site also presents Coding Mode and Work Mode side by side as delivery-oriented product forms; details are as given on each product page.

Current model versions (public pages, 2026-10-25).

  • Text: MiniMax-M3.1-Flash (first row of the open platform's PAYG price table; its Hugging Face model card gives about 428B total parameters, about 23B activated, up to about 1M context, and text/image/video input; the benchmark tables on the card are vendor tests) and MiniMax-M3 (PAYG prices split by ≤512k / >512k input and by standard / priority tier; its model card likewise gives about 428B / 23B and 1M context as vendor claims). The M Plan docs note that MiniMax-M3.1-Flash-Preview has been retired and requests using the old name are served by MiniMax-M3.1-Flash.
  • Video: the flagship MiniMax-H3 and variants such as Highspeed / Lite / Max, billed per output second and resolution; 768P video produced by H3 can also be regenerated in 2K (billed separately). Legacy Hailuo 2.3 / 02 are still listed under "legacy models".
  • Speech: speech-2.8-hd / speech-2.8-turbo (sync and async), plus ASR, voice design and fast cloning.
  • Music: the open platform says that from 2026-08-20 the paid endpoints (music generation, lyrics generation) are no longer offered to new users (existing paid users can keep using them), and the free music endpoints (Music-3.0-free, Music-2.6-free, music-cover-free) have been shut down; users are pointed to MiniMax Audio or the Music 3 open weights on Hugging Face / ModelScope.

Keep M Plan and PAYG apart. M Plan is a personal subscription (Go / Explore / Build) whose quota is for MiniMax Code and compatible tools; subscription keys and normal PAYG API keys cannot be mixed (FAQ wording). All three tiers include M3.1 Flash and image / audio models; Go excludes H3 video, while Explore / Build include H3. For production, the FAQ recommends pay-as-you-go.

Intelligence with Everyone

MiniMax About page mission line

Core capabilities

Text: a fast model and a flagship sold in tiers. On PAYG prices, M3.1-Flash costs far less per unit than M3; M3 is further tiered by whether input exceeds 512k and by standard / priority (set service_tier=priority in the call; the docs say about 1.5× the standard price). Cache reads are priced separately. Agent / coding / cybersecurity benchmark figures on the model cards and blog are all marked as vendor tests, and the scaffolds used for each benchmark (such as Claude Code and Codex) are as noted on the model cards.

Video: H3 per second, packages don't cover H3. The H3 series is billed by output resolution and second; input images have a free allowance, then a per-image fee; input video is billed by duration; audio input is free (per the flagship table). The pricing page states: video packages support the Hailuo series and don't yet support MiniMax H3, so H3 goes through PAYG or a custom business deal. The international site has the same structure, priced in USD. Hailuo packages deduct "video points", with different deductions by resolution and duration, and failed or safety-blocked tasks deduct no points.

Speech: PAYG plus subscriptions / packages. China bills in units such as "CNY per 10,000 characters" (rules such as one Chinese character counting as 2 characters are on the pricing page); international TTS bills in "$ per M characters". The international site offers an Audio Subscription (Starter–Business monthly table, with quarterly / yearly discounts); China has voice packages instead (HD / Turbo series, with validity, character allowance and RPM listed in the pricing overview). Voice design and fast cloning are billed "on first use in synthesis" (per the official notes), not when the voice_id is created.

Open weights and licenses. The Hugging Face LICENSE for both MiniMax-M3 and MiniMax-M3.1-Flash is the MiniMax Community License: non-commercial use is free; commercial use must prominently display "Built with MiniMax M3" or "Built with MiniMax M3.1-Flash"; if the related products or services earn more than USD 20M a year, prior written authorization is required (the email subject format is in the LICENSE), otherwise a one-time notice email is enough. An appendix lists prohibited uses (including military purposes and harming minors). The product apps and hosted API are outside that license's open-source scope.

H3 weights, separate license. The official self-hosting page says H3-Base (FL2VA, Ref2VA) can be downloaded from Hugging Face MiniMaxAI/MiniMax-H3, and the open release does not include H3-Context-IR or 2K regeneration. H3 uses its own MiniMax H3 Community License Agreement (licensor Nanonoble Pte. Ltd., license date 2026-08-02), not the M3 Community License. Its Applicable Territory is "worldwide, except the Excluded Territories", which are defined as the European Union, the United Kingdom, South Korea and the United States; use in those territories requires a separate license from MiniMax. Commercial terms: prior written authorization if the licensee's commercial products and services earn more than USD 20M a year; commercial products using H3 must prominently display "MiniMax H3" in the interface; the acceptable-use policy lists prohibited uses (including military use). The territory clause applies to the H3 weight license; the open platform's pricing pages state no territory limit.

Other open weights. The self-hosting overview also lists open weights for MiniMax-M2.7 and MiniMax Music 3 and notes that licenses differ by model: the M2.7 LICENSE is non-commercial, free for personal self-hosting, non-profit and academic research, with commercial use requiring prior written authorization; Music 3 uses the MiniMax-Music3 Community License, under which commercial products must display "MiniMax-Music3" in the interface, and prior written authorization is required when the related products and services provided by the licensee and its affiliates earn more than USD 20M a year in total.

M Plan quota windows. FAQ: non-video capabilities are limited by a 5-hour window and a weekly window; video only by the weekly window; quota doesn't roll over. Both windows start from first use and are independent of each other. Annual billing "charges only 10 months". Credit packs can be bought separately; the China FAQ gives 1,000 credits = ¥7, converted at par with PAYG list prices, valid for 1 year from purchase; plan quota is used first, then credit packs.

Server tools and MCP. The pricing overview lists per-call prices for web_search, API-vlm and others; when called through M Plan they deduct plan quota or credits as the docs describe. Whether a tool is included in a tier is set out in the M Plan model table and the console; a subscription doesn't mean unlimited use of every open-platform tool.

How it compares

Subscription (M Plan) vs PAYG API. M Plan suits interactive personal use in MiniMax Code or compatible IDEs / agents; the FAQ says it is for individuals and recommends PAYG for production. A subscription key deducts only plan quota / credit packs and never automatically draws on account balance; entering a normal API key by mistake does draw on the balance, which the docs cover in a separate FAQ entry. The Go tier can't generate H3 video; video-heavy workflows suit the Explore / Build tiers or PAYG H3.

Open weights vs hosted. Self-hosting M3 / M3.1-Flash avoids per-token fees, but you must follow the Community License's attribution and high-revenue authorization terms and bring your own inference resources. The hosted API offers capabilities such as Speech and web tools that have no matching weight repos; self-hosted H3-Base excludes Context-IR and 2K regeneration and is bound by the H3 license's territory terms.

China site vs international site. platform.minimax.cn prices in CNY; platform.minimax.io prices in USD. Unit prices for the same model name can't be compared directly at the exchange rate. M Plan relative quota: the China FAQ and China M Plan overview both put Build at 12× Go, while the international overview puts Build at 7.5× Go; the two sides' docs disagree, and each account follows the docs of its own site.

Video billing forms. H3 is the per-second flagship route; legacy Hailuo models can use video packages. If a workflow is tied to H3, buying Hailuo packages saves nothing on H3. 2K regeneration charges the output and input material fees again, on top of the original task.

Versus single-line vendors. Compared with vendors that sell only video or only coding subscriptions, MiniMax puts both in one account system: the upside is one platform for multimodal work, the cost is more billing concepts (PAYG / M Plan / packages / credit packs). Most of the selection effort goes into understanding the key types rather than comparing any single unit price.

Who it's for / Not for

Who it's for

Product teams that need to call text agents, video generation and speech synthesis from one vendor (open-platform PAYG).

Developers who work mainly in MiniMax Code and accept personal subscription quota windows (M Plan Go / Explore / Build; video needs Explore or above).

Individual developers who want to attach subscription quota to several coding tools and can accept "multiple tools sharing the same quota" (the FAQ says simultaneous use is allowed, but the quota is shared).

Teams able to run their own inference and accept the MiniMax Community License's commercial terms (downloading the M3 or M3.1-Flash weights).

Creators or studios that need the H3 flagship for video work and accept per-second PAYG or a custom business deal.

Who it's not for

Teams that require the consumer clients or hosted API to be open-source and auditable end to end: the platform is a closed commercial service.

People using M Plan as an unlimited production API: the official FAQ explicitly recommends PAYG for production, and there are RPM/TPM limits plus the two windows.

People who want "video packages" alone to cover H3: the pricing page says packages don't yet support H3.

Teams planning to use M3 / M3.1-Flash in commercial products or services whose yearly revenue may exceed USD 20M without arranging prior written authorization.

Teams that want to self-host H3 weights in the Excluded Territories listed in the H3 license (EU, UK, South Korea, US) without a separate license.

New users who still want to open the open platform's paid music API: the vendor says it is no longer offered to new users from 2026-08-20.

Getting started

Open platform PAYG. China accounts are on platform.minimax.cn and need a top-up; international accounts are on platform.minimax.io. PAYG calls use a normal API key, with model IDs listed in the pricing overview (such as MiniMax-M3.1-Flash, MiniMax-H3, speech-2.8-hd). The priority tier requires setting service_tier in the request as the docs describe. Accounts can set a low-balance alert (pricing FAQ).

M Plan. Buy Go / Explore / Build following the "subscribe" doc; sign in to MiniMax Code and it works, with no need to configure a key. When connecting tools such as Claude Code / Cursor / Codex, use the subscription key; entering a PAYG key draws down the account balance. Annual billing charges 10 months' fees, per the official notes. The FAQ says downgrades aren't supported yet; after turning off auto-renew, the current cycle stays usable. The China FAQ says mmx auth login detects the service region from the API key; if 401 persists after login, set region=cn for a China key (mmx config set --key region --value cn) or global for an international key.

Self-hosting. hf download MiniMaxAI/MiniMax-M3 or MiniMaxAI/MiniMax-M3.1-Flash, read the corresponding LICENSE, and deploy with SGLang / vLLM or other frameworks recommended on the model card. Before commercial use, complete the attribution wording and (where applicable) the authorization email. Deploy H3-Base following the official H3 self-hosting page, and read the H3 license's territory terms before downloading.

Voice / video packages. Buy only after confirming the model series matches (voice packages split into HD / Turbo; video packages cover Hailuo only). Validity and the no-rollover rule for remaining balance are in the pricing overview.

PAYG unit prices, whether packages still "exclude H3", and whether M Plan promotions are still running can all change on the day of ordering; figures are as of 2026-10-25.

Pricing & license

As of 2026-10-25. API and resource-package figures come from the open platform's pricing overviews. Intl M Plan regular monthly prices come from the Regular price column of the official monthly-offer page (the half-price first-month promo ended 2026-10-14).

CNY PAYG (selected). Flash ¥1.00/¥4.00/cache ¥0.10 per MTok. M3 standard ≤512k, listed as permanent 50% off (list ¥4.20/¥16.80/¥0.84): ¥2.10/¥8.40/cache ¥0.42. H3 ¥0.50/s at 768P, ¥0.80/s at 2K (input video billed at the same tier; images beyond the free allowance charged separately). speech-2.8-hd ¥3.50/10k chars, turbo ¥2.00/10k chars. Image image-01 ¥0.025/image. web_search ¥0.03/call. Tiered list and discounted prices, and validity periods, for the speech and Hailuo video resource packages are in the full table in the pricing overview and are not copied here; before buying, check that the package matches the model series.

USD PAYG (selected). Flash $0.15/$0.60/cache $0.015. M3 standard ≤512k, listed as permanent 50% off (list $0.60/$2.40/$0.12): $0.30/$1.20/$0.06. H3 $0.08/s 768P, $0.13/s 2K. speech-2.8-hd $100/M chars, turbo $60/M chars. Audio Subscription Starter $5–Business $999/mo (quarterly/yearly discount tables). Video Packages start at about $1,000 for a 1-month Standard package (Hailuo only).

Intl M Plan regular. Go $22, Explore $55, Build $132 /mo. Relative quota: the international docs say Build = 7.5×Go; the CN FAQ says Build = 12×Go (they disagree; see caveats). Annual billing charges 10 months (FAQ).

License. Hosted: closed commercial. Weights: MiniMax Community License for M3 and M3.1-Flash (attribution wording required; prior authorization above the $20M revenue threshold); MiniMax H3 Community License (with Excluded Territories) for H3; separate licenses for M2.7 and Music 3.

Risks & limits

The two key types don't mix. Subscription keys and PAYG API keys are separate; the FAQ's example of "still charged balance after subscribing" usually means a PAYG key was entered in the tool. Details such as the order in which credit packs and plan quota are used, and that credit packs can be bought without M Plan but carry no Code membership benefits, are as stated in the FAQ.

Quota windows and no rollover. The M Plan 5-hour and weekly windows, the weekly-only window for video, and quota reset at the end of each billing cycle can drain an allowance that looks large during a burst of video jobs. Video isn't limited by the 5-hour window but is by the weekly one, so waiting for the 5-hour refresh doesn't restore video weekly quota.

H3 and packages don't match. Video packages don't cover H3; before a custom deal, check whether the quote is per second or in package points. Regenerating a video in 2K bills the original input material again.

License appendices. The Community License prohibits uses such as military purposes; the H3 license adds an Applicable Territory clause (excluding the EU, UK, South Korea and the US) that applies only to the weight license. Get a legal review before commercial deployment.

Regional docs disagree. The Build-to-Go multiplier differs between the CN FAQ and the international overview.

Preview model name. If code still uses MiniMax-M3.1-Flash-Preview, the docs say it maps to the released Flash, but switch to the current ID in long-lived configs to avoid misreading issues during troubleshooting.

Vendor figures. User counts, country coverage and model-card benchmarks are vendor claims and vendor tests.

Music API closed to new customers. The policy in effect since 2026-08-20 is a hard limit for new integrations; use the Audio product or the open weights instead.

Our verdict

MiniMax's difference is "multimodal under one roof": text agents, H3 video and Speech can run on the same open platform, with M Plan keeping personal coding subscriptions separate from PAYG production traffic. The choice comes down to three factors: whether H3 is needed (which decides whether video packages apply and whether M Plan must be at least Explore), whether the account is mainland or international (currency and the M Plan multiplier docs), and whether the weights are self-hosted (whether the Community License applies). Treating the subscription as a production API, or the open weights as freely usable MIT code, runs into limits the docs already spell out.

Chat-only use does not necessarily suit it; teams already using Hailuo or Speech that plan to add agentic coding are a better fit, since one vendor cuts the cost of integrating multiple accounts, provided someone reads the exception clauses in the pricing overview.

Why it matters

MiniMax builds its own models for text, video, speech and music, and offers its products to users worldwide.

Sources

Official

  1. MiniMax - About Us | MiniMaxminimax.io · Accessed Oct 25, 2026
  2. MiniMax Investor Relationsir.minimax.io · Accessed Oct 25, 2026
  3. MiniMaxminimax.io · Accessed Oct 25, 2026
  4. Pricing Overviewplatform.minimax.io · Accessed Oct 25, 2026
  5. M Plan overview - MiniMax API Docsplatform.minimax.io · Accessed Oct 25, 2026
  6. FAQ - MiniMax API Docsplatform.minimax.io · Accessed Oct 25, 2026
  7. M Plan overviewplatform.minimax.io · Accessed Oct 25, 2026
  8. M Plan monthly: 50% off your first monthplatform.minimax.io · Accessed Oct 25, 2026

Updates & reading

Long readEditorial

A Tougher AI Leaderboard Leaves More Than Open-Weight Models Behind

Artificial Analysis says its Intelligence Index v5 widened the gap between open-weight models and the leader to 12 points from 3, and critics say a test that is 60% private can't be checked. The firm's own per-test scores show that many closed models fell alongside open-weight ones on the new tests, and that more than half of DeepSeek's deficit to the top models comes from public benchmarks anyone can rerun.