Back to glossary

Term

MiniMax

MiniMax is an AI lab founded in Shanghai in 2021 whose open-weight M model family (LLMs) targets very long context windows and low running costs via sparse attention.

MiniMax — explained in detail

MiniMax is an AI lab based in Shanghai, founded in late 2021 by former employees of the computer-vision company SenseTime. It develops its own foundation models and positions itself among the Chinese providers that release their models as open weight — comparable to DeepSeek, Qwen (Alibaba) or Moonshot/Kimi.

The term refers both to the company and to its M model family — a series of large language models (LLMs). Earlier generations were branded abab; since 2025 MiniMax has numbered the models as M1, M2.x and M3. Alongside the language models, MiniMax also runs consumer products for video and music generation under the Hailuo brand.

What “open weight” means here

Open weight means the trained model weights are published, so the model can be downloaded and run on your own hardware — unlike pure API models such as GPT or Claude, whose weights stay closed. Open weight is not the same as fully open source: training data and the complete training code usually remain undisclosed. In practice, there can also be a gap between the announcement and the actual release of the weights.

The technical distinctive: sparse attention + long context

The defining approach of the more recent M models is sparse attention (MiniMax calls its variant MiniMax Sparse Attention, MSA). Classic attention computes every token against all previous tokens; the compute cost therefore scales poorly with context length. Sparse attention instead evaluates only a subset of these connections, sharply reducing the per-token compute cost at long context.

This is the lever behind the MiniMax models: they reach very large context windows — up to around one million tokens — while staying comparatively cheap to run. Rather than simply making the model bigger, MiniMax shifts the optimisation toward the efficiency of long contexts. That is especially relevant for agentic workflows, which carry a lot of context anyway.

Example / Practical use

The family’s current model, MiniMax M3, serves as an example: introduced in June 2026, it combines sparse attention with around one million tokens of context and native multimodality and targets agentic coding. MiniMax advertises coding benchmarks at frontier level for a fraction of the cost of closed providers — though at launch these figures came from the provider itself, with independent tests and the release of the weights partly still pending.

→ Details, benchmark figures and context for that specific release are in the news article MiniMax M3: Open-Weight Frontier Model with Sparse Attention.

The glossary entry itself describes MiniMax as a provider and model family — it is not narrowed to a single model version, since M3 will be followed by further generations.

  • MiniMax vs. individual models (M3, M2.x): MiniMax is the provider and the family; M3 is one concrete version of it. Model versions age, the provider remains.
  • MiniMax vs. DeepSeek / Qwen / Kimi: All four are Chinese open-weight providers pushing on price and openness at the same time. They differ in architecture, focus areas (coding, reasoning, multimodality) and licensing details.
  • Sparse attention vs. mixture of experts (MoE): Both cut compute cost, but in different places — sparse attention reduces the attention computation across the context, MoE activates only part of the model parameters per token. MiniMax combines both approaches.
  • MiniMax (AI provider) vs. minimax (algorithm): The name overlaps with the classic minimax algorithm from game theory / AI search — here it refers exclusively to the company and its model family.

Entdecke mehr

Themenuebersicht