As an AI consultancy from Rosenheim, we support companies across the DACH region (Germany, Austria, Switzerland) with GDPR-compliant integration of open-weight models such as MiniMax-M3. Through self-hosting in EU data centres, the model can be deployed for agentic workflows and long-context tasks in compliance with data protection regulations.
MiniMax-M3: Frontier Coding with 1M Context
MiniMax unveiled MiniMax-M3 on 1 June 2026 – according to the lab, the first open-weight model to combine frontier coding, a 1-million-token context window and native multimodality (image and video understanding) in a single model.
MSA Architecture (MiniMax Sparse Attention)
The new architecture design is the key to its efficiency on long contexts:
- Up to 20x lower per-token compute at 1M context compared to the predecessor
- More than 9x faster prefill, more than 15x faster decoding
- Significantly lower inference cost for long-context workloads
Native Multimodality
M3 understands text, images and video in a unified model – ideal for agentic coding with UI screenshots, document processing and multimodal research.
Coding Performance
- 59.0% on SWE-Bench Pro – beats GPT-5.5 and Gemini 3.1 Pro, narrowly behind Claude Opus 4.7
- Optimised for complex, multi-step software engineering tasks
- Well-suited for autonomous agent harnesses
API Pricing
- approx. $0.60 / 1M input tokens and $2.40 / 1M output tokens
- Launch promotion: 50% off ($0.30 / $1.20)
- Substantially cheaper than Claude Opus 4.7 or GPT-5.5 at comparable coding performance
Open Weights
MiniMax has announced that weights will be released on Hugging Face and GitHub within approximately 10 days of launch – enabling self-hosting on your own EU infrastructure.
Previous Generation: MiniMax-M2.7
MiniMax-M2.7 (released May 2026) remains relevant for pure coding-agent workflows without multimodality and with smaller context requirements. It offers:
- Sparse Mixture-of-Experts (MoE) architecture
- 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, 66.6% on MLE-Bench Lite
- MIT licence, fully open source
For new projects, however, we recommend going directly with MiniMax-M3 thanks to its 1M context, multimodality and higher coding performance.
EU Deployment Options
Self-Hosting in EU Data Centres
For GDPR compliance, we offer support with:
- Deployment on AWS EU regions (Frankfurt, Ireland)
- Azure EU regions (West Europe, Germany)
- Google Cloud EU regions (Frankfurt, Belgium)
- Private cloud or on-premise in your own data centre
Hardware Requirements
MiniMax-M3 and M2.7 are available in various quantisations:
- BF16: Full precision for research and benchmarks
- FP8: Recommended for production deployment
- INT4/INT8: Efficient quantisation for limited resources
Alternative API Access
For rapid prototyping without your own infrastructure:
- MiniMax API (platform.minimaxi.com): Official access from MiniMax AI
- Third-party providers: Together.ai, Fireworks AI, OpenRouter
- Model weights: Hugging Face (MiniMaxAI) and GitHub
Note: Direct API usage occurs via Chinese infrastructure and is not GDPR-compliant without a data processing agreement and self-hosting.
Local or Cloud? Calculate the Cost
For agentic coding workloads, a precise cost calculation between API usage and self-hosting often pays off. The local-vs-cloud AI inference calculator from ai-prices.eu lets you compare hardware, electricity and operating costs against cloud API pricing and determine the break-even point for your workload.
Integration & Support
Our Recommendation
Self-hosting in EU data centres is the best option for GDPR-compliant usage of MiniMax-M3. We support you with:
- Infrastructure planning and hardware sizing (especially for 1M-context workloads)
- Deployment, quantisation and optimisation
- Integration into existing agent frameworks
- Compliant usage within CompanyGPT
- Fine-tuning for specific use cases
For companies seeking a leading open-weight model with frontier coding, 1M context and native multimodality, MiniMax-M3 is an excellent choice – provided the appropriate infrastructure is available.