Tracked as a source-level fact for technical comparison.
Source-verified model profile · 1 October 2026
MiniMax M2.5
MiniMax M2.5 is an open-weight model from MiniMax focused on coding, search, tool use, agentic workflows and office work. This profile separates model facts, license terms, infrastructure feasibility and European data-residency questions.
OPEN WEIGHTSMODIFIED MITASIAEU SELF-HOSTING POSSIBLE
Direct answer
What is MiniMax M2.5?
MiniMax M2.5 is a MiniMax M2.5 open-weight model published by MiniMax. The current profile records 229B-class, sparse moe, 196K maximum per individual sequence in the publisher vLLM deployment guide of context and Text/code → text/code. The recorded license is Modified MIT.
Its practical focus is coding, search, tool use, agentic workflows and office work. Treat the exact checkpoint and revision as the deployable entity; quantization and serving layers can materially change memory, throughput and exposed features.
Official model source ↗Provider origin🇨🇳 China
RegionAsia
Parameters229B-class
Active computeSparse MoE
Context196K maximum per individual sequence in the publisher vLLM deployment guide
ModalitiesText/code → text/code
LicenseModified MIT
VerificationSource-verified · 2026-10-01
Architecture & capability
Publisher-documented signals.
Tracked as a source-level fact for technical comparison.
Tracked as a source-level fact for technical comparison.
Tracked as a source-level fact for technical comparison.
Tracked as a source-level fact for technical comparison.
License & commercial use
Separate downloadability from rights.
The recorded license is Modified MIT. Open weights describe availability of parameters; they do not automatically settle commercial use, redistribution, derivative works, attribution, patent terms or a separate acceptable-use policy.
For production, retain the exact license/revision alongside the deployed checkpoint and confirm whether provider-specific conditions apply.
Technical comparison only — not legal advice.
Deployment
Hardware and runtime reality.
Datacenter / multi-GPU class; publisher deployment guidance cites hundreds of GB of weight/KV-cache memory.
MiniMax documents SGLang, vLLM, Transformers and KTransformers deployment paths.
Context length, batch size, KV cache, quantization, expert routing and multimodal encoders can shift memory and throughput substantially.
EU deployment lens
Provider origin is not data residency.
EU/EEA self-hosting is technically possible when the downloadable weights, runtime and connected services are operated on infrastructure selected by the deployer.
The 🇨🇳 flag describes provider origin: China. It does not say where prompts, documents, embeddings, logs, monitoring or backups are processed.
Self-hosting does not itself establish GDPR or AI Act compliance. Map the complete data flow and the organization's legal role.
EU deployment & data-residency guide →FAQ
Common questions.
What is MiniMax M2.5?
MiniMax M2.5 is an open-weight model from MiniMax focused on coding, search, tool use, agentic workflows and office work.
Can MiniMax M2.5 be self-hosted?
Yes, the published weights enable independent deployment where the license permits it. Hardware feasibility depends on precision, context and serving topology.
Can MiniMax M2.5 run in the EU?
Technically yes when the complete inference stack and connected data services are operated on EU/EEA infrastructure.
Can MiniMax M2.5 be used commercially?
The recorded license is Modified MIT. Review the exact official terms and any separate use policy before production.
Primary source
Verify the exact checkpoint.
The official model card remains authoritative for revision-specific architecture, files, inference settings and license links.
MiniMax · official model page ↗