MiniMax M3: Specs, Pricing, Context Window and Availability Facts

MiniMax M3 (API ID MiniMax-M3) is an open-weight, natively multimodal mixture-of-experts language model from MiniMax, released on 1 June 2026, which MiniMax's models page describes as a "Frontier multimodal coding model with 1M context window". This profile gathers the published facts about it in one place: parameter counts, context window, prices for both service tiers, the thinking controls, API parameters, open-weight licence terms, provider-reported benchmarks and deprecation status. Every fact comes from MiniMax's own documentation, its Hugging Face model card, its licence file or its research paper, listed under Standards and references. Where MiniMax publishes nothing on a point, the page says so.

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 documentation as of November 2026. Model details change, so confirm against MiniMax's current documentation.

Key takeaways

• Released 1 June 2026. MiniMax's model release notes record MiniMax M3 on that date as the latest M-series language model for agentic reasoning, tool use, coding, multimodal chat input and long-context tasks.
• About 428B total, 23B active, 1M tokens of context. The Hugging Face card states roughly 428B total and 23B active parameters, and MiniMax's API documentation states a 1,000,000-token context window.
• $0.30 input and $1.20 output per million tokens. These are Standard rates up to 512K input tokens, marked "Permanent 50% off"; rates double above 512K, and the Priority tier costs 1.5x Standard.
• Open weights under the MiniMax Community License. Commercial use requires a "Built with MiniMax M3" notice, a one-time notice to MiniMax up to USD 20 million yearly revenue, and prior written authorisation above it.
• No deprecation date. MiniMax's release notes list no retirement notice for M3, and the text generation guide says it "also remains available" alongside the newer MiniMax-M3.1-Flash.

What is MiniMax M3?

MiniMax M3 is the third generation of MiniMax's M-series text models, which began with MiniMax-M2 on 27 October 2025. It is the first in the series that MiniMax describes as natively multimodal: the model card states that M3 "undergoes mixed-modality training from the very first step", and it accepts text, image and video input and returns text.

Its main architectural change is MiniMax Sparse Attention (MSA), which replaces standard grouped-query attention for long sequences. The model card reports 9x prefill and 15x decode speed-ups over M2 at 1M context, with per-token compute reduced to 1/20. MiniMax published the attention method in an arXiv paper, "MiniMax Sparse Attention" (2606.13392, first submitted 11 June 2026), and the inference kernel on GitHub.

M3 is also a reasoning model with three thinking modes. In MiniMax's coding agent, MiniMax Code, and in the API, it targets software engineering and agentic work. For the full MiniMax line-up, see MiniMax Models. For how sparse expert models work, see Mixture of Experts Explained.

Key facts at a glance

As of 2026-11-15. Sources: MiniMax models page, API reference, release notes, Hugging Face model card and licence.

Release and version history

As of 2026-11-15. Dates from MiniMax's model release notes and the arXiv record.

MiniMax lists one API identifier, MiniMax-M3, with no dated snapshots and no "-highspeed" variant on the pages checked. The model card notes LHTB 2.0 results added in October 2026, so the card itself has been updated since launch; check it for later changes.

How much does MiniMax M3 cost?

As of 2026-11-15. USD per million tokens. Source: MiniMax pay-as-you-go pricing page.

• Discount wording: each M3 rate carries the label "Permanent 50% off"; the page gives no effective date. Budget against the listed price and recheck the page, since pricing labels can change.
• Long-context threshold: every rate doubles once a request exceeds 512K input tokens, so the effective price changes halfway through the 1M window.
• Priority: set service_tier to priority; the API reference says Priority gets faster admission at 1.5x the Standard price.
• Thinking tokens: the API reference states that thinking tokens count toward max_tokens, so they are billed as output.

Worked example at Standard rates: a request with 100,000 input tokens and 10,000 output tokens costs $0.03 for input plus $0.012 for output, $0.042 in total. The same request at Priority costs $0.063. For how token billing and caching work across providers, see Model Pricing Explained and the LLM Pricing Tracker.

Context window and output limits

MiniMax states a 1,000,000-token context window for M3 in its API reference. It does not state a separate maximum output figure on the pages checked; output is bounded by the max_tokens value the caller sets, which includes thinking tokens. Self-hosted context depends on the serving setup and GPU memory.

A count_tokens endpoint (POST /anthropic/v1/messages/count_tokens) estimates input tokens without generating output, which helps keep requests under the 512K price boundary. See Context Windows Explained.

Modalities and tools

M3 accepts text, image and video and returns text. MiniMax's API reference describes an Anthropic-compatible endpoint at https://api.minimax.io/anthropic, and the text generation guide also lists an OpenAI-compatible endpoint for MiniMax language models.

• Thinking: the model card lists three modes, enabled, adaptive and disabled. In the Anthropic-compatible API thinking is off by default; set type adaptive to turn it on. Thinking blocks must be passed back unchanged in later turns, especially in tool-use conversations.
• Tool use: tools are supported; tool_choice accepts auto only. In multi-turn tool calls, append the full response content (thinking, text and tool_use blocks) to the message history.
• Fixed sampling: temperature is ignored and fixed at 1; top_p is ignored and fixed at 0.95. The model card recommends the same values for self-hosting.
• Ignored parameters: top_k, stop_sequences, mcp_servers, context_management and container are ignored. output_config.effort applies only to MiniMax-M3.1-Flash, not M3.

Tools that reach into code repositories or shells give the model credentialed access to systems. Scope each connected tool to least privilege; see the LLM Security Guide.

Where is MiniMax M3 available?

As of 2026-11-15. Sources: Hugging Face model card and MiniMax API documentation.

Third-party hosts and community quantisations are run by their own operators, under their own data terms, so test any quantised build before relying on it. See Self-Hosted vs Hosted LLM APIs.

Access restrictions

M3 is not a restricted-access model: it is listed on MiniMax's public pricing page and its weights are downloadable from Hugging Face. The restrictions that apply sit in the licence rather than an access programme. The MiniMax Community License grants rights for non-commercial purposes and allows commercial use only on conditions:

• Attribution: prominently display "Built with MiniMax M3" on a related website, interface, blog post, about page or product documentation, and keep the copyright and permission notice in all copies.
• Up to USD 20 million yearly revenue: send a one-time notice to MiniMax by email.
• Over USD 20 million yearly revenue: obtain separate prior written authorisation from MiniMax.
• What counts as commercial: paid products or services that rely on the software, commercial use of APIs for it or its derivatives, and deploying post-trained or fine-tuned versions for a commercial purpose.
• Use restrictions: no unlawful content, no military purpose, no harm to minors, no harmful false information and no promotion of discrimination or hate.

The licence is MiniMax's own, not an OSI-approved open source licence; the file does not state a governing law. Have counsel read the full text before commercial deployment. See Open-Weight vs Closed LLMs Explained and the Open-Weight Model Licence Tracker.

Benchmarks (provider-reported, dated)

As of 2026-11-15. All figures are provider-reported, from the MiniMax-M3 Hugging Face model card (launched June 2026; LHTB 2.0 added October 2026). The card's benchmark chart is an image and names comparators only there, so no comparator column is given. Independent leaderboard results are not included because none were checked against their own published pages for this profile.

Higher is better on all six. Harness, scaffold and time budget change agentic scores, so compare only results run the same way. See How to Read LLM Benchmarks and the LLM Leaderboard Snapshot.

Model card and system card

MiniMax publishes a model card on Hugging Face and a research paper on the attention method. It does not publish a separate system card for M3 on the pages checked. What the published material covers:

• Model card: parameter counts, architecture (MSA, mixture of experts), native multimodal training, thinking modes, benchmark scores, recommended sampling settings, serving frameworks and quantisation links.
• Research paper: the MSA design, an index branch that selects a Top-k subset of key-value blocks per grouped-query attention group and a main branch running exact block-sparse attention on them, with kernel co-design for GPUs. Reported speed-ups of 14.2x prefill and 7.6x decode on H800 GPUs apply to the paper's 109B research model.
• Safety evaluations: Not stated by MiniMax on the pages checked (2026-11-15); the model card has no safety or limitations section.
• Training data and knowledge cutoff: Not stated by MiniMax on the pages checked (2026-11-15).

Teams that need documented risk evaluations should record this gap in their model assessment and run their own red-team tests. See Model Cards and System Cards Explained.

Data handling

• Hosted API training use and retention: Not stated by MiniMax on the pages checked (2026-11-15). Read MiniMax's privacy policy and platform terms before sending personal or confidential data.
• Zero data retention: Not stated by MiniMax on the pages checked (2026-11-15).
• Self-hosted weights: running the published weights keeps prompts and outputs on the operator's own infrastructure, and the operator owns access control, logging and patching.
• Files API: large videos are uploaded to MiniMax's Files API and referenced by file ID, so they are stored on MiniMax's side; retention for uploaded files is not stated on the pages checked.
• Jurisdiction: confirm where the hosted service processes data and which law applies; where Chinese rules are relevant, see the guide below. See the China Generative AI Measures and AI Labelling Rules Guide.

MiniMax API keys are non-human identities with spend and data exposure. See the LLMjacking Guide, How to set spend limits on LLM API keys and the Self-Hosted LLM Security Guide.

Deprecation status

As of 15 November 2026, MiniMax's model release notes contain no deprecation or retirement notice for MiniMax-M3, and the text generation guide says M3 "also remains available" after the release of M3.1-Flash. MiniMax does not publish a dedicated LLM deprecation schedule on the pages checked.

Watch the release notes and models page for changes. See the LLM Deprecation and Retirement Tracker.

How MiniMax M3 differs from other models in its family

As of 2026-11-15. Sources: MiniMax release notes, models page, API reference and Hugging Face cards. Models listed by release date.

The documented differences: M3 was the first natively multimodal M-series model and introduced MSA. M3.1-Flash keeps M3's mixture-of-experts design but, according to its card, replaces the first three full-attention layers with sparse attention, applies quantisation-aware training to the routed experts and attention paths, swaps the earlier EAGLE-like multi-token prediction for DSpark speculative decoding, and adds tunable thinking depth through output_config.effort, which M3 ignores. The M3.1-Flash card text and its Hugging Face size metadata give different parameter counts (428B and 220B), so confirm the size before capacity planning. See MiniMax Models for prices and details of the other models.

Practitioner checklist

• Pin MiniMax-M3 explicitly and record the date adopted; there are no dated snapshots, so retest when the model card changes.
• Set max_tokens with thinking tokens in mind, and pass thinking blocks back unchanged in multi-turn tool use.
• Remove temperature, top_p, top_k and stop_sequences logic that the API ignores, and use tool_choice auto only.
• Use count_tokens to keep requests under 512K input tokens where cost matters, and set per-key spend limits before enabling Priority.
• For self-hosting, check the commercial conditions of the MiniMax Community License, send the required notice or obtain authorisation, and display the attribution.
• Benchmark M3 and M3.1-Flash on your own workload with the same harness before choosing; vendor scores use specific scaffolds.
• Read MiniMax's privacy policy and platform terms, and decide whether regulated data goes to the hosted API or stays on self-hosted weights.
• Store MiniMax API keys in a secrets manager, separate them by environment and rotate them on staff changes.
• Watch the MiniMax release notes monthly for changes to M3.

Standards and references

• MiniMax: MiniMax-M3 model card on Hugging Face (accessed 2026-11-15)
• MiniMax: Models introduction (accessed 2026-11-15)
• MiniMax: Anthropic API compatible reference (accessed 2026-11-15)
• MiniMax: Text generation guide (accessed 2026-11-15)
• MiniMax: Pay-as-you-go pricing (accessed 2026-11-15)
• MiniMax: Model release notes (accessed 2026-11-15)
• MiniMax: MiniMax Community License (M3) (accessed 2026-11-15)
• arXiv: MiniMax Sparse Attention (2606.13392) (accessed 2026-11-15)
• MiniMax: MiniMax-M3.1-Flash model card on Hugging Face (accessed 2026-11-15)

Related NHI Mgmt Group resources: MiniMax Models, Open-Weight Model Security Guide, Frontier AI Models Compared and LLM Release Tracker.