MiniMax Unveils M3, a Domestic AI Large Model That Surpasses GPT-5.5
China’s AI sector has witnessed a major technological leap with the official launch of Xiyu Technology’s latest large language model, MiniMax M3. This advanced system combines state-of-the-art coding proficiency with support for an ultra-long context window of up to 1 million tokens. Notably, it natively supports multi-modal inputs, including image and video processing, alongside computer desktop interaction, making it the first open-source model in China to integrate these three critical capabilities.

Exceptional Results in Standardized Benchmarks
In the rigorous SWE-Bench Pro programming evaluation, MiniMax M3 secured a score of 59.0%, outperforming GPT-5.5 and Gemini 3.1 Pro, while closely trailing the top-tier Opus 4.7. The model also demonstrated superior performance in the Claw-Eval test for AI agent capabilities and achieved leading scores on the OmniDocBench multi-modal document understanding benchmark.
Advanced Architecture Enhances Efficiency
Driving this performance is M3’s adoption of a novel sparse attention architecture (MSA). When handling ultra-long contexts of 1 million tokens, this design reduces single-token computation by half compared to its predecessor. Consequently, the model accelerates comprehension speed by more than 9 times and boosts answer generation speed by over 15 times. The API is now live, with the official team committing to release model weights and technical reports for global developers within 10 days.
Related article
India mandates caller-ID apps to share spam data with telecom operators
India has expanded its anti-spam regulations to mandate that caller-ID and call-management applications share user spam reports with telecom operators, a move that has led Truecaller, a prominent spam-blocking app provider, to label the policy as ant
Qwen AI Platform Expands Model Matrix With Official Launch of GLM-5.3 and DeepSeek-V4-Pro
The Alibaba Cloud Qwen AI platform (MaaS) has recently expanded its model service matrix, integrating Zhipu’s flagship large model GLM-5.3 and the official release of DeepSeek-V4-Pro. The corresponding API services are now publicly available, enablin
Why Agents Forget and Go Off-Track in Long Tasks: AWS, Claude Code, and Manus Unpack Four Frameworks
Large language models frequently lose focus during extended operations, drifting away from their original objectives—a persistent challenge in the intelligent agent sector. The issue often stems not from the model itself, but from the surrounding exe
Related Special Topic Recommendations
Comments (0)
0/500
China’s AI sector has witnessed a major technological leap with the official launch of Xiyu Technology’s latest large language model, MiniMax M3. This advanced system combines state-of-the-art coding proficiency with support for an ultra-long context window of up to 1 million tokens. Notably, it natively supports multi-modal inputs, including image and video processing, alongside computer desktop interaction, making it the first open-source model in China to integrate these three critical capabilities.

Exceptional Results in Standardized Benchmarks
In the rigorous SWE-Bench Pro programming evaluation, MiniMax M3 secured a score of 59.0%, outperforming GPT-5.5 and Gemini 3.1 Pro, while closely trailing the top-tier Opus 4.7. The model also demonstrated superior performance in the Claw-Eval test for AI agent capabilities and achieved leading scores on the OmniDocBench multi-modal document understanding benchmark.
Advanced Architecture Enhances Efficiency
Driving this performance is M3’s adoption of a novel sparse attention architecture (MSA). When handling ultra-long contexts of 1 million tokens, this design reduces single-token computation by half compared to its predecessor. Consequently, the model accelerates comprehension speed by more than 9 times and boosts answer generation speed by over 15 times. The API is now live, with the official team committing to release model weights and technical reports for global developers within 10 days.
India mandates caller-ID apps to share spam data with telecom operators
India has expanded its anti-spam regulations to mandate that caller-ID and call-management applications share user spam reports with telecom operators, a move that has led Truecaller, a prominent spam-blocking app provider, to label the policy as ant
Qwen AI Platform Expands Model Matrix With Official Launch of GLM-5.3 and DeepSeek-V4-Pro
The Alibaba Cloud Qwen AI platform (MaaS) has recently expanded its model service matrix, integrating Zhipu’s flagship large model GLM-5.3 and the official release of DeepSeek-V4-Pro. The corresponding API services are now publicly available, enablin
Why Agents Forget and Go Off-Track in Long Tasks: AWS, Claude Code, and Manus Unpack Four Frameworks
Large language models frequently lose focus during extended operations, drifting away from their original objectives—a persistent challenge in the intelligent agent sector. The issue often stems not from the model itself, but from the surrounding exe





Home






