MiniMax (M3 / M2.7): making long context cheap with sparse attention, and the licence question for schools

Short answer: MiniMax has taken the "use sparse attention to make long context cheap" route, the engineering is coherent and the prices really are low — but a bespoke weights licence and a mainland API mean a Hong Kong school has to clear legal and privacy questions before technical evaluation even starts.

What this is

MiniMax is a Shanghai model team whose product line spans language, speech, music and video models. On the language side its positioning has been consistent: not the largest model, but the same long context at lower cost.

That positioning goes back to MiniMax-Text-01, an early attempt at ultra-long context using linear-attention variants that influenced later long-context designs. By 2026, M3 uses the team's own MSA sparse attention instead. Schools do not need the algorithmic detail, only the conclusion: for the same 1M-token context window, MiniMax's per-token cost is below most of its peers.

The current line-up

• MiniMax M3 (around June 2026): about 428B total parameters with roughly 23B active per token, MSA sparse attention and a 1M-token context window. Promotional pricing is $0.30 per 1M input tokens and $1.20 per 1M output tokens, against a list price of $0.60 / $2.40. The weights are downloadable from Hugging Face under MiniMax's own licence.
• MiniMax M2.7 (open-sourced on 12 April 2026): the release that established the agent-oriented, low-cost positioning, still common across deployments.
• MiniMax-Text-01 (January 2025): an early ultra-long-context attempt using linear-attention variants, with a genuine place in the technical history.

The full version and price comparison is in the automatically generated series table further down this page.

Strengths

1. Low unit cost at long context. This is the concrete selling point. Where a workflow repeatedly reads large document sets — a whole year group's coursework, or large volumes of meeting records — the cost gap shows clearly on the bill.
2. Sparse attention is a real engineering contribution. Moving 1M-token context from "possible but expensive" to "possible and affordable" is one reason long context became ordinary in 2026, and it deserves technical credit.
3. Downloadable weights. An organisation able to host them controls the data path completely instead of depending on a cloud API — provided the licence permits it and the hardware exists.
4. Efficient inference from the MoE design. With only about 23B parameters active per token, latency and throughput beat a dense model of the same total size, which matters for interactive use.

Limits

• The licence is bespoke. It is neither MIT nor Apache 2.0, and the terms need reading clause by clause. A school or sponsoring body putting it into a school system should get a legal opinion rather than rely on the word "open".
• The API is in mainland China. Calling it directly is a cross-border data transfer, and anything touching staff or pupil personal data requires a Personal Data (Privacy) Ordinance assessment first.
• Promotional pricing is not long-term pricing. The list price is double the promotional rate. Multi-year budgets should be built on list price, treating the promotion as headroom rather than as the baseline.
• Self-hosting remains demanding. All 428B parameters must fit in memory, needing hardware well beyond an ordinary school server room. The organisations that benefit here are the ones already running GPU clusters.
• Almost no education material exists. No Hong Kong curriculum prompt templates, classroom case studies or teacher community, so an IT coordinator works it out alone and support costs run above the mainstream providers.

How a Hong Kong school should view it

The practical stance is to treat MiniMax as a way of understanding why long context got cheap, rather than as a procurement candidate.

A school that still wants to evaluate it should work in this order. First the licence: give the Hugging Face terms to the sponsoring body's legal adviser and confirm that commercial use, redistribution and attribution requirements fit school use. Second, data residency: if the plan is to call the official API, complete the cross-border transfer assessment honestly and explain to parents how pupil data is handled; if the plan is to self-host, that gate disappears and is replaced by a realistic hardware and maintenance assessment. Only third comes educational suitability — Chinese quality, Traditional-character usage and content-policy boundaries, all tested with the school's own questions rather than read off a benchmark.

It is worth adding that "cheap" is often overrated in a school context. Under the Edor.ai model, a school pays a fixed annual fee with AI usage included and no token bill, so a vendor's per-token price is not a direct school expense at all. Where data is processed, and how student use is gated, deserve much more attention.

Availability inside Edor.ai

Not wired up today. The platform natively supports OpenAI, Azure OpenAI, Anthropic, Poe and local Ollama; MiniMax is not among them and administrators will not see the option. The reasons are the two above: bespoke licence terms that need clause-by-clause legal review, and an official API in mainland China that constitutes a cross-border transfer.

We have no partnership with MiniMax. This page simply sets out verifiable information so a school can judge for itself.

Alternatives

• For an open model that runs on a single machine with data kept on site, see the 27B class under Alibaba Qwen.
• For permissive, standard licensing, see the Apache 2.0 weights under Tencent Hunyuan or the MIT-licensed releases under Zhipu GLM.
• To compare the Chinese models on price, licensing and availability, read the DeepSeek, Kimi, Qwen and GLM comparison.
• To understand the difference between long context and retrieval, start with the relevant lessons in the LLM Classroom.