MiniMax Models: Every Model, Spec, Price and Release Date

MiniMax is an AI company that offers its M-series large language models through the MiniMax API Platform and publishes their weights on Hugging Face, and it released four M-series generations between October 2025 and October 2026. This hub lists every MiniMax text model in MiniMax's public API documentation, release notes and Hugging Face model cards as of 15 November 2026, with API IDs, release dates, context windows, prices, licences, availability and benchmark figures. It covers MiniMax-M3.1-Flash, MiniMax-M3, the M2.7 and M2.5 series and the earlier M2, M2.1 and Text-01 models. It is written for developers, buyers, architects and anyone who needs to know which MiniMax model is which. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from MiniMax's own documentation or its official Hugging Face cards, cited in the sources list.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses MiniMax's public API documentation as of November 2026. Model details change, so confirm against MiniMax's current documentation.

Key takeaways

• Two text models are current on the API. MiniMax's models page lists MiniMax-M3.1-Flash and MiniMax-M3; the M2.7 and M2.5 series are not listed on the current models or pricing pages.
• Both current models list 1M context. MiniMax-M3 and MiniMax-M3.1-Flash each list a 1M-token context window; maximum output is not stated.
• Prices start at $0.15 per million input tokens. MiniMax-M3.1-Flash costs $0.15 input and $0.60 output; MiniMax-M3 costs $0.30 and $1.20 up to 512K input tokens after a permanent 50% discount.
• Weights are published under a community licence. The M3 and M3.1-Flash cards name the MiniMax Community License, which requires attribution and prior authorisation for commercial products above $20 million yearly revenue.
• No retirement dates are published. MiniMax's release notes and models page show no text-model deprecation or discontinuation notices.

What is MiniMax?

MiniMax develops the M-series large language models alongside video (Hailuo and H3), image, speech and music models. Its text models are offered through the MiniMax API Platform, its MiniMax Agent product and a Coding Plan subscription, and their weights are published on the MiniMaxAI organisation on Hugging Face. The M-series began with MiniMax-M2 in October 2025; MiniMax-M3 in June 2026 added native image and video input and MiniMax Sparse Attention for 1M-token contexts.

The main table lists the M-series text models named in this page's scope plus the newer M3.1-Flash. Non-text models are summarised under families and naming. Facts are as of 15 November 2026.

Current MiniMax models at a glance

As of 2026-11-15. Source: MiniMax models page, pricing page, release notes and Hugging Face cards. Models are listed by release date, newest first. Category is the MiniMax model line.

MiniMax describes both current models as multimodal frontier coding models, with M3.1-Flash adding adjustable thinking depth. "Not listed" records what the current pages show; MiniMax has not published a retirement notice for these models. See Open-Weight vs Closed Models Explained.

Key facts for the main MiniMax models

As of 2026-11-15. Identical fields for every model; "Not stated by MiniMax" means not stated by MiniMax on the pages checked (2026-11-15). Parameter counts are those on the Hugging Face model card or size box.

The two M3.1-Flash parameter figures conflict on the same Hugging Face page, and MiniMax does not explain the difference; check the repository before sizing hardware. See Mixture of Experts Explained.

Features stated across current models

• Sparse attention: MiniMax-M3 uses MiniMax Sparse Attention (MSA) instead of standard grouped-query attention, with reported 9x prefill and 15x decode speed-ups over M2 at 1M context (M3 card).
• Thinking controls: M3 offers a thinking parameter with enabled, adaptive and disabled modes; M3.1-Flash uses a reasoning_effort setting with low, medium, high, xhigh and max levels.
• Native multimodality: the M3 card states the model was trained on mixed modalities from the first step.
• Recommended sampling: temperature 1.0 and top_p 0.95 on M3; temperature 1.0, top_p 0.95 and top_k 40 on M2.5 and M2.7.
• Service tiers: M3 has Standard and Priority tiers, selected with service_tier set to priority at 1.5 times the Standard price.
• Server tools: web_search and an API-vlm MCP tool are priced per request on the pricing page.

MiniMax model families and naming

• M3.1 Flash: the newest text model, released on 19 October 2026 as an open-weight model, with adjustable reasoning effort.
• M3: released on 1 June 2026, MiniMax's first natively multimodal M-series model, with 1M context and Standard and Priority service tiers.
• M2.7 series: MiniMax-M2.7 and MiniMax-M2.7-highspeed, released on 18 March 2026 for the API Platform and MiniMax Agent.
• M2.5 series: MiniMax-M2.5 and MiniMax-M2.5-highspeed, released in February 2026; the announcement names the faster variant M2.5-Lightning at a steady 100 tokens per second against 50 for the standard model, with identical capability.
• M2 and M2.1: MiniMax-M2 (27 October 2025) and MiniMax-M2.1 (22 December 2025).
• Earlier text models: MiniMax-Text-01 and MiniMax-VL-01, released on 15 January 2025.
• Non-text models: the H3 video family (H3, H3-Highspeed, H3-Lite, H3-Max), H3 Image, Hailuo video models, Speech-2.8 text-to-speech and Music-3.0. Music generation and lyrics APIs closed to new users on 20 August 2026.
• Suffixes: "-highspeed" marks a faster-serving variant of the same model; "Flash" marks the efficiency-focused M3.1 model. See LLM Naming and Versioning Explained.

When was each MiniMax model released?

As of 2026-11-15. Dates are from the MiniMax model release notes and announcement pages, in date order. Non-text releases are omitted except where noted.

The latest text model release is MiniMax M3.1 Flash on 19 October 2026. Between M2 and M3.1-Flash, MiniMax released a new M-series text model roughly every two to four months, and the release notes list each M2-era model as a series with a highspeed variant from M2.5 onwards. See the LLM Release Tracker.

How much do MiniMax models cost?

As of 2026-11-15. Price per million tokens in USD. MiniMax does not list a batch rate or cache-write price on the pricing page checked; the columns are kept for parity with other provider hubs. Source: MiniMax pay-as-you-go pricing.

• Undiscounted M3 rates: the page shows the earlier Standard rates of $0.60 input, $2.40 output and $0.12 cached input up to 512K, and $1.20, $4.80 and $0.24 above 512K.
• Long-context tier: M3 prices double once a request exceeds 512K input tokens.
• Non-text models: H3 video models are priced per second of output by resolution; MiniMax-H3-Image costs $2.30 per million input text tokens, $1.50 per million input image tokens and $3.80 per million output image tokens; speech-to-text is $0.38 per hour; text-to-speech is $100 per million characters for speech-2.8-hd and $60 for speech-2.8-turbo; voice design is $3 per voice and rapid voice cloning $1.50 per voice.
• Token size: the pricing page states that roughly 750 English words equal 1,000 tokens.
• Tool charges: web_search and API-vlm (MCP) cost $0.01 per request each.
• Subscription: MiniMax also sells a Coding Plan subscription that includes M-series model access.

See Model Pricing Explained and How to set spend limits on LLM API keys.

Where are MiniMax models available?

• First-party: the MiniMax API Platform, MiniMax Agent, MiniMax Code and the Coding Plan subscription.
• Open weights: the MiniMaxAI organisation on Hugging Face hosts the M3.1-Flash, M3, M2.7 and M2.5 weights; quantised GGUF builds for llama.cpp, Ollama and LM Studio are listed for M3.
• Third-party hosting named by MiniMax: the M3 card lists Hugging Face Inference Providers and HuggingChat.
• Serving frameworks named on the M3 card: SGLang, vLLM, Hugging Face Transformers, KTransformers, Unsloth, ATOM (ROCm) and Docker Model Runner.
• Cloud platforms and regions: major cloud availability and processing regions are not stated by MiniMax on the pages checked (2026-11-15).

Self-hosting moves operation, patching and access control to the organisation. See the Self-Hosted LLM Security Guide.

Restricted-access MiniMax models

As of 2026-11-15. Source: MiniMax models page, pricing page and Hugging Face cards.

No MiniMax text model is described as restricted to verified or invited organisations on the pages checked.

What is the context window of MiniMax models?

MiniMax-M3 and MiniMax-M3.1-Flash each list a 1M-token context window on the models page, and the M3.1-Flash card notes that usable length depends on GPU memory and serving configuration. M3 pricing changes at 512K input tokens. Context windows for M2.7, M2.5, M2.1 and M2 are not stated on the pages checked (2026-11-15), and no MiniMax page checked states a maximum output length. See Context Windows Explained.

Is any MiniMax model open source?

MiniMax publishes weights for its recent M-series models, under licences that differ by model. The M3 and M3.1-Flash cards name the MiniMax Community License. Its base grant covers non-commercial use; commercial use, including paid services, commercial API use and deploying fine-tuned versions, requires displaying "Built with MiniMax M3", a one-time notice to MiniMax below $20 million yearly revenue and separate prior written authorisation above it. Its appendix prohibits uses including military purposes and harm to minors. The M2.5 card names a Modified-MIT licence, and the M2.7 card links a licence file tagged "Other". Training data is not published. See the Open-Weight Model Security Guide.

Model cards and system cards

As of 2026-11-15.

The M3 card has no safety section. See Model Cards and System Cards Explained.

Benchmarks

As of 2026-11-15. Provider-reported and independent results are kept in separate columns. Scores are as MiniMax reports them on the Hugging Face cards.

The M2.5 card's evaluation metadata gives SWE-bench Verified as 75.8 against 80.2% in its announcement text, without explanation. Benchmark versions differ between cards, so scores are not directly comparable across models. See How to Read LLM Benchmarks.

Deprecations and retirements

As of 2026-11-15. Source: MiniMax models page and release notes. MiniMax does not publish a fixed notice period on the pages checked.

Code that still calls an M2-series API ID should be tested against the current API before relying on it. See the LLM Deprecation and Retirement Tracker.

Data handling and security

• Privacy policy: the MiniMax API Platform links a privacy policy at platform.minimax.io; retention, training-use and storage-location terms are not stated by MiniMax on the pages checked (2026-11-15).
• Zero data retention: not stated by MiniMax on the pages checked (2026-11-15).
• Self-hosted use: running the published weights keeps prompts and outputs in the operator's environment.
• Licence obligations: commercial self-hosting of M3 or M3.1-Flash carries attribution, notice and revenue-threshold conditions.
• Key handling: generic controls apply to every provider's API: keep keys out of code, scope and rotate them, and set spend limits.

See AI Data Retention and Deletion and the Foundation Model Selection Guide.

How to choose a MiniMax model: factual criteria

• Price per token: M3.1-Flash is listed at half the M3 Standard input rate; M3 doubles above 512K input tokens.
• Latency and priority: M3's Priority tier costs 1.5 times Standard.
• Reasoning control: M3 offers enabled, adaptive and disabled thinking; M3.1-Flash offers five reasoning_effort levels.
• Modalities: M3 and M3.1-Flash accept text, image and video; M2.5 and M2.7 are text models.
• Hosting model: API use means data is processed under MiniMax's terms; self-hosting the weights keeps data in-house.
• Hardware for self-hosting: M3 has about 428B total parameters; M2.5 and M2.7 list 229B.
• Licence: the Community License for M3 and M3.1-Flash, Modified-MIT for M2.5; check revenue thresholds before commercial use.
• Policy constraints: check internal and public sector rules on the use of the provider before adoption.

Practitioner checklist

• Move any MiniMax-M2, M2.1, M2.5 or M2.7 API calls to a model listed on the current models page and test the output.
• Record the exact model ID and service tier in the AI inventory.
• For commercial products on M3 weights, send the licence notice or obtain authorisation, and add the required attribution.
• Confirm the M3.1-Flash parameter count from the repository files before sizing GPUs.
• Assess data residency, retention and training-use terms against internal policy before sending regulated data to the API.
• Download weights only from the official MiniMaxAI Hugging Face organisation and track provenance.
• Store MiniMax API keys in a secrets manager and set spend limits. See the LLMjacking Guide.

Standards and references

• MiniMax: Models (accessed 2026-11-15)
• MiniMax: Pay-as-you-go pricing (accessed 2026-11-15)
• MiniMax: Model release notes (accessed 2026-11-15)
• MiniMax: MiniMax-M3 model card (accessed 2026-11-15)
• MiniMax: MiniMax-M3.1-Flash model card (accessed 2026-11-15)
• MiniMax: MiniMax Community License (accessed 2026-11-15)
• MiniMax: MiniMax-M2.7 model card (accessed 2026-11-15)
• MiniMax: MiniMax-M2.5 model card (accessed 2026-11-15)
• MiniMax: MiniMax M2.7 announcement (accessed 2026-11-15)
• MiniMax: MiniMax M2.5 announcement (accessed 2026-11-15)

Related NHI Mgmt Group resources: DeepSeek Models, Self-Hosted LLMs vs Hosted LLM APIs, Open-Weight Model Security Guide, LLM Security Guide.