China AI Native Industry Insights – 20250618 – MiniMax | Moonshot AI | more

Explore MiniMax’s groundbreaking M1, the world’s first open-source hybrid architecture inference model, Hailuo 02’s record-breaking video model efficiency, and Moonshot AI’s Kimi-Dev-72B, a leading open-source coding LLM for bug fixing. Discover more in Today’s China AI Native Industry Insights.

1. MiniMax Debuts M1: World’s First Open-Source Hybrid Architecture Inference Model

🔑 Key Details:
– Million-Token Context: M1 supports 1M token inputs (matching Google Gemini 2.5 Pro) and 80K token outputs, exceeding most open-source models.
– Hybrid Architecture: Novel lightning attention mechanism enables efficient long-context processing using just 30% of comparable model computing resources.
– Advanced Performance: M1 excels in SWE-bench (56%), long-context understanding (global #2), and tool use scenarios (outperforming Gemini 2.5 Pro).
– Cost-Effective Training: Reinforcement learning completed in just three weeks using 512 H800 GPUs, costing only $534,700.

💡 How It Helps:
– AI Researchers: Access to an efficient architecture that dramatically reduces computational requirements for long-context models.
– Software Developers: Superior performance in software engineering tasks with 56% on SWE-bench validation benchmark.
– Enterprise Users: Cost-effective API pricing with tiered rates based on context length, starting at ¥0.8/million tokens.
– Data Analysts: Process and analyze documents up to 1M tokens in length with competitive understanding capabilities.

🌟 Why It Matters:
MiniMax M1 sets a new standard for cost-efficient, long-context AI. Its hybrid design challenges the notion that top-tier performance demands vast compute, offering both accessibility and scalability. This opens the door for smaller teams and organizations to adopt high-capacity models for deep analysis, research, and enterprise applications.

Original Chinese article: https://mp.weixin.qq.com/s/OWrbRE3zHaeNahXkPnP0fg

English translation via free online service: https://translate.google.com/translate?hl=en&sl=zh-CN&tl=en&u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%2FOWrbRE3zHaeNahXkPnP0fg

Video Credit: MiniMax (official) (@MiniMax__AI on X)

2. Hailuo 02: Breaking Global Records in Video Model Efficiency

🔑 Key Details:
– Breakthrough Architecture: Hailuo 02 uses innovative Noise-aware Compute Redistribution (NCR) architecture, improving training efficiency by 2.5x.
– Enhanced Performance: 3x larger parameter count and 4x more training data than previous version, enabling superior instruction following and complex physics rendering.
– Native 1080p: First model capable of generating affordable native 1080p videos with state-of-the-art instruction compliance.
– Competitive Pricing: Offers industry-leading prices compared to both domestic and international competitors.

💡 How It Helps:
– Content Creators: Enables generation of complex scenes like gymnastics that were previously impossible with AI video models.
– Video Artists: Supports higher resolution (1080p) output while maintaining cost-effectiveness.
– Prompt Engineers: Better instruction following allows more precise control over generated content.
– Budget-Conscious Users: Provides premium video generation capabilities at lower cost than competitors.

🌟 Why It Matters:
Hailuo 02 represents a significant advancement in democratizing video generation technology, aligning with the company’s mission of “Intelligence with Everyone.” By dramatically improving efficiency without increasing user costs, this release redefines the price-performance ratio in AI video generation. Its ability to handle previously impossible scenes positions it as a potential market leader, while setting the stage for future improvements in generation speed, preference alignment, and advanced functionality.

Original Chinese article: https://mp.weixin.qq.com/s/hrBG4J1eSnANlaxhlY5UbQ

English translation via free online service: https://translate.google.com/translate?hl=en&sl=zh-CN&tl=en&u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%2FhrBG4J1eSnANlaxhlY5UbQ

Video Credit: Hailuo AI (MiniMax) (@Hailuo_AI on X)

3. Moonshot AI Releases Kimi-Dev-72B: Leading Open-Source Coding LLM for Bug Fixing

🔑 Key Details:
– SOTA Performance: Achieves 60.4% on SWE-bench Verified, setting new benchmark among open-source models.
– RL Optimization: Uses Docker-based reinforcement learning with rewards for passing test suites.
– Dual Architecture: Implements both BugFixer and TestWriter capabilities through a two-stage process.
– Open Access: Available on Hugging Face and GitHub for community use and development.

💡 How It Helps:
– Software Engineers: Autonomous resolution of complex coding issues with test-verified solutions.
– Open-Source Contributors: Accessible model with scaling capabilities for integration into development workflows.
– Research Teams: Framework for combined bug-fixing and test-writing through self-play mechanism.

🌟 Why It Matters:
Kimi-Dev-72B represents a significant advancement in open-source AI for software development, bridging the gap between human-like reasoning and automated code fixes. Its release democratizes access to high-performance coding assistance while pioneering reinforcement learning methodologies that could influence future model training approaches.

Original article: https://moonshotai.github.io/Kimi-Dev/

Video Credit: The original article

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.