What Is MiniMax? Language, Speech and Video Model Lines
MiniMax is a company brand spanning several model lines. This article focuses on general language and coding models, distinguishing the M3 series, earlier M2.5 weights, and the roles of Speech, Music and H3.
Meet this model
MiniMax is an AI company and model brand. Its models cover language and coding, with separate speech, music and video generation lines. “Using MiniMax to code” generally requires a specific M-series language model, which is distinct from the MiniMax Agent app or a Speech model. The exact model determines input types and thinking controls, while an older weight license does not replace current hosted service terms.
Choose language and media generation models separately
The official October 20, 2026 catalog lists MiniMax-M3.1-Flash and M3 as language models with one-million-token contexts and multimodal input. H3 is listed for video generation and Speech for voice. Access to a provider’s API does not mean that one model ID can handle every media task.
To fix a web page from an interface screenshot, confirm that the selected M model accepts images, then supply project files through a coding tool. To create narration for a tutorial, choose a speech model separately. A language model writing narration text does not automatically produce audio. This distinction prevents a model mismatch from being mistaken for an insufficiently long prompt.
Compatible interfaces retain their own parameter rules
The official Anthropic-compatible interface states that M3.1-Flash always thinks and returns 400 when thinking.type is disabled. M3 has thinking off by default and can enable it with adaptive. The interface also ignores some parameters, including temperature, so copying another service’s sampling settings does not establish a comparison under identical conditions.
For multi-turn tool calls, the documentation requires retaining the full assistant response, including thinking and tool blocks, before continuing with tool results. Keeping only the final text can lose information needed for the next call. When a continuing task fails, check protocol-compliant conversation handling instead of merely switching to a supposedly stronger model name.
Read the license shipped with open weights
The official MiniMax-M2.5 model card still provides local deployment instructions, but its capabilities are not the current M3 service specification. The card is labeled modified-mit, while the accompanying LICENSE-MODEL is titled MiniMax Model License and includes requirements such as supplying the agreement and retaining a Notice when distributing. Read the release’s license file rather than inferring standard MIT terms from a page label.
Local deployment instructions list inference frameworks such as SGLang and vLLM; tool parsing and hardware must still match the release. This article does not run MiniMax models or verify provider coding scores or response speeds. For selection, start with a real failing code example and compare whether a fix preserves existing behavior before migrating an entire project.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.