MiniMax M3 Just Killed Closed-Source Models

2026-06-04 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Software Development & Engineering · Depth: Advanced, quick

Summary

MiniMax shipped its M3 model on June 1, 2026, positioning it as a frontier-class, open-weights coding model with a 1M-token context window. Notably, M3 is priced at approximately 5–10% of rivals like GPT-5.5 and Gemini 3.1 Pro. This cost efficiency stems from a novel approach to attention called MiniMax Sparse Attention (MSA), which challenges the conventional need to process the entire past context for next-token prediction in long sequences. The author notes that most benchmarks are MiniMax's own, and while the model's weights are not yet available for independent verification, the analysis focuses on the mechanism and reported numbers.

Key takeaway

For Machine Learning Engineers evaluating long-context models, MiniMax M3's introduction signals a significant shift towards more cost-effective open-weights solutions. You should investigate sparse attention architectures for your own model development, especially when aiming for large context windows without prohibitive inference costs. Consider M3 as a benchmark for future open-source coding models, and prepare to test its performance once weights become available.

Key insights

MiniMax M3 demonstrates sparse attention enables cost-effective, long-context, open-weights coding models.

Principles

• Long context doesn't require full past attention.
• Sparse attention improves inference speed and cost.

In practice

• Achieve frontier coding at 5-10% cost.
• Utilize 1M-token context windows.
• Explore open-weights alternatives.

Topics

• MiniMax M3
• Sparse Attention
• Open-weights Models
• Coding Models
• Long Context
• AI Inference Cost

Best for: CTO, VP of Engineering/Data, AI Engineer, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps — Full-attention LLMs are intrinsically sparse and can be efficiently transformed into sparse models with minimal training.
• MiniMax Releases M2.1: An Enhanced M2 Version with Features like Multi-Coding Language Support, API Integration, and Improved Tools for Structured Coding — MiniMax M2.1 enhances open-source coding and automation with multi-language support and cost-efficiency.
• Z.ai GLM-5: New SOTA Open Weights LLM — Chinese AI labs are rapidly advancing open-weight models, emphasizing cost-efficiency, long-context, and agentic capabilities.
• Can Large Language Models Understand Context? — LLMs struggle with nuanced context understanding, especially pre-trained dense models and those with aggressive quantization.
• SparseBalance: Load-Balanced Long Context Training with Dynamic Sparse Attention — SparseBalance co-optimizes sparse attention and sequence length heterogeneity for efficient LLM training.
• Gemini 3.5 Pro X-High, MiniMax M3, DeepSwe, New Claude Models, MiMO-v2.5 Upgrade, & More! AI NEWS — Major AI model updates, architectural innovations, and commercial deployments are rapidly advancing AI capabilities and applications.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.