MiniMax has officially released the M2.1 large language model, featuring a sparse MoE architecture, a 200K context window, and 10B activated parameters. It excels at multilingual programming, complex task handling, and efficient reasoning, supporting various mainstream programming languages and Web3 protocols. Benchmarks like SWE-bench Multilingual and VIBE show it outperforms closed-source models such as Claude Sonnet 4.5 and Gemini 3 Pro. The API is now live and supports local deployment, priced at 8% of Claude 4.5 Sonnet.

Zhipu AI has released its new-generation flagship large language model GLM-4.7, focusing on stability in engineering scenarios, long-task processing, and multilingual programming. It demonstrates excellent performance on benchmarks including SWE-bench, LiveCodeBench, and Terminal Bench, with overall capabilities surpassing closed-source models like GPT-5.2. Supporting a 200K context, it is now the default model for GLM Coding Plan. The open-source weights are available on Hugging Face.

OpenAI's latest FrontierScience evaluation shows GPT-5.2 scored 92% on the 'Google-Proof' science benchmark, significantly exceeding the 70% expert baseline. It also leads competitors like Claude Opus 4.5 and Gemini 3 Pro in Olympiad-style reasoning (77%) and complex open-ended research tasks (25%), preliminarily validating its ability to accelerate scientific research.

OpenAI has introduced the Skills feature in Codex. Based on the open Agent Skills specification led by Anthropic, it allows developers to extend AI coding agent capabilities via preset packages or custom scripts defined with natural language. Interoperability has been achieved with multiple platforms including Microsoft, GitHub, and Cursor, promoting standardization in the AI development tool ecosystem.

Google has officially released the Gemini 3 Flash model, focusing on ultra-high speed, low cost, and multimodal processing (text, image, audio, video). Its reasoning speed is 3 times that of the 2.5 Pro model, with substantially lower API pricing. It is now the default model for products like the Gemini App and AI Search, suitable for high-frequency real-time scenarios.

In 2025, the AI industry's focus shifted from a parameter race to application innovation and industrial adoption. Open-source large models like DeepSeek, GLM-4.7, and MiniMax M2.1 caught up with or even surpassed closed-source models on multiple benchmarks. Agentic AI, standardized protocols (e.g., MCP, Agent Skills), multimodal generation, and hardware infrastructure optimization became key annual trends. Enterprise focus moved from single performance metrics to practical application and ROI.

Alibaba's Tongyi Qwen3-TTS family has launched the voice creation model Qwen3-TTS-VD-Flash and the voice cloning model Qwen3-TTS-VC-Flash. They support complex natural language instructions for customizing voice, 3-second-level voice cloning, and generation in 10 mainstream languages. Overall performance surpasses models like GPT-4o-mini-tts and MiniMax, significantly enhancing multilingual TTS capabilities.

Zoho has launched its self-developed Zia LLM suite, focusing on enterprise finance automation and privacy protection. It uses multi-specification models with 1.3B-7B parameters, integrates the MCP protocol, and supports over 700 business actions. Performance surpasses Llama 3-8B in tasks like structured data extraction. Plans to scale to 100B parameters by year-end aim to promote a new standard for 'Vertical AI' and automated finance.

Nvidia announced an acquisition of all assets (excluding cloud business) of AI chip startup Groq for approximately $20 billion. Groq's founding team will join Nvidia. Known for its TPU architecture and low-latency inference chips, this marks Nvidia's largest acquisition ever, further solidifying its dominance in the AI inference chip market.

In 2025, the AI developer ecosystem focuses on standardized infrastructure like Agentic CLI, MCP protocol, and Agent Skills. Open-source models like GLM-4.7, MiniMax M2.1, and DeepSeek V3.2 are catching up to closed-source in engineering, long-context, and multimodal capabilities. Enterprise focus has shifted to practical application, ROI, and multi-scenario deployment, with AI engineering and industrial synergy becoming mainstream trends.