# The New China AI Trifecta

Source: https://www.youtube.com/watch?v=82DyXL0ZXI8
Recap page: https://rapidrecap.app/video/82DyXL0ZXI8
Generated: 2026-01-13T16:10:31.727+00:00

---
## Quick Overview

The modern Chinese AI landscape is characterized by a competitive trifecta of open-source labs—Moonshot AI, Z.ai (formerly Zhipu AI), and MiniMax—who are rapidly advancing AI capabilities, often surpassing or matching proprietary models across various benchmarks while focusing heavily on practical application, agentic capabilities, and efficient inference via techniques like Grouped-Query Attention and Quantization-Aware Training, leading to significant industry shifts and high valuations.

**Key Points:**
- Moonshot AI's Kimi K2 achieved first place in the video generation benchmark, outperforming competitors like DeepSeek and Qwen 1.5, despite being much smaller.
- Z.ai (formerly Zhipu AI) achieved a high valuation of $5.6B in February 2025 and its GLM-4.7 model ranked first in the Agentic Tool Use benchmark (96%) and second in IFBench for instruction following (72%), indicating strong agentic capabilities.
- MiniMax released MiniMax-M1 and M2, with M2 achieving 61.0% on SWE-bench Verified, surpassing DeepSeek v3.2 (60.0%) and GLM-4.6 (55.4%) at a significantly lower cost ($214 vs $14 for DeepSeek).
- The shift in AI development prioritizes practicality, agentic coding, tool use, and code reasoning over pure reasoning and raw knowledge, as evidenced by the focus of GLM-4.5 Air training.
- MiniMax's M2 model uses techniques like Lightning Attention to achieve high performance and long context windows (up to 4M tokens for inference), while Z.ai's Kimi K2 uses Quantization-Aware Training to deliver high performance in a smaller package (355B vs 1T parameters for Kimi K2).
- The Chinese AI ecosystem is highly competitive, with companies like Z.ai (HKEX: 02513) achieving public listing and rapid growth, placing intense pressure on US giants like Anthropic and OpenAI.
- MiniMax's M2 model was an experimental model internally named M2-mini, which unexpectedly performed so well during pretraining that the company decided to release it formally.

![Screenshot at 00:09: The video introduces the concept of open-source AI releases from different labs between July and December, setting the stage for comparing their various models and philosophies.](https://ss.rapidrecap.app/screens/82DyXL0ZXI8/00-00-09.jpg)

**Context:** The video analyzes the rapid advancements and competitive landscape of the Chinese AI industry, focusing on three key open-source labs: Moonshot AI (creators of Kimi), Z.ai (creators of GLM models), and MiniMax. The context highlights a shift from focusing solely on raw knowledge benchmarks to prioritizing practical application, agentic capabilities (like tool use and coding), and inference efficiency. This competition is placing pressure on established Western leaders like OpenAI and Anthropic, evidenced by recent high valuations and benchmark performances of the Chinese models.

## Detailed Analysis

The video details the rapid advancements in the Chinese open-source AI sector, highlighting a trifecta of competitive labs: Moonshot AI (Kimi), Z.ai (GLM), and MiniMax. The overall trend shows a pivot from purely academic benchmarks to practical, agentic capabilities. Moonshot AI's Kimi K2, despite being much smaller than competitors like LLaMA-3 70B, outperformed them on several benchmarks (02:17). Z.ai, which recently went public (HKEX: 02513) (09:09), emphasizes agentic capabilities, with GLM-4.7 leading on the Agentic Tool Use benchmark (96%) (08:00) and utilizing Grouped-Query Attention (07:08). MiniMax has also made significant strides, releasing MiniMax-M1 and M2. M2 achieved the top spot on SWE-bench Verified at 61.0% score for only $214 in inference cost, outperforming competitors like DeepSeek and GLM-4.6 (11:29). The video stresses that this progress is achieved through novel techniques: Moonshot uses a whale-inspired architecture, Z.ai uses an optimization technique called Muon to replace AdamW (03:36), and MiniMax employs Lightning Attention and MoE layers to handle massive context windows (up to 4M tokens for inference) efficiently (10:05). Furthermore, the competitive valuations are noted: Z.ai is valued at $5.6B (06:04), Moonshot AI received funding that valued it at $4B (12:02), and MiniMax also achieved a $4B valuation (11:57). The increasing focus on application-driven AI (like coding and tool use) over pure research is presented as a key dynamic in this new AI era, exemplified by GLM-4.6V's native multimodal tool use capabilities (08:47).

### Chinese Open Source Trifecta

- Moonshot AI (Kimi) leads in video generation benchmarks
- Z.ai (GLM) excels in agentic tool use with GLM-4.7
- MiniMax (M2) tops SWE-bench at high efficiency

### Key Technological Innovations

- Kimi K2 uses a smaller model size with high performance
- Z.ai uses Muon optimizer to replace AdamW
- MiniMax uses Lightning Attention for massive context windows (1M tokens training, 4M inference)

### Performance Metrics

- GLM-4.7 scores 96% on Agentic Tool Use; MiniMax M2 scores 61% on SWE-bench Verified for $214 cost
- GLM-4.5 (Air) focuses on Agentic formatting and Practicality over pure reasoning benchmarks

### Industry Dynamics & Valuation

- Z.ai reached a $5.6B valuation
- Moonshot AI and MiniMax valued around $4B
- Chinese labs are increasingly outperforming US counterparts in specific areas like agentic tasks.

### MiniMax M2 Development

- Was internally named M2-mini and was an experimental model that performed surprisingly well during pretraining, leading to its formal release.

### Benchmarking Shift

- The industry is moving from pure reasoning/raw knowledge benchmarks (MMLU, GSM8K) to practical, agentic benchmarks (SWE-bench, Context-Bench, IFBench) (02:03, 07:55).

### Openness vs. Proprietary

- Moonshot AI and MiniMax released their models under the permissive MIT license, contrasting with the closed nature of some proprietary models.

![Screenshot at 00:01: The title card shows the DeepSeek V3.2 paper abstract, highlighting its high computational efficiency and agentic capabilities.](https://ss.rapidrecap.app/screens/82DyXL0ZXI8/00-00-01.jpg)
![Screenshot at 00:42: A comparison of three logos representing the three key players in the Chinese AI trifecta: Moonshot AI \(piano keys\), Z.ai \(Z logo\), and MiniMax \(waveform\).](https://ss.rapidrecap.app/screens/82DyXL0ZXI8/00-00-42.jpg)
![Screenshot at 02:07: A slide contrasting 'THEN' \(Reasoning, Raw Knowledge\) with 'NOW' \(Practicality\), illustrating the shift in AI focus.](https://ss.rapidrecap.app/screens/82DyXL0ZXI8/00-02-07.jpg)
![Screenshot at 06:05: A bar chart comparing models on the Artificial Analysis Intelligence Index, showing MiniMax M2 \(Quantized\) scoring 67, competitive with proprietary models.](https://ss.rapidrecap.app/screens/82DyXL0ZXI8/00-06-05.jpg)
![Screenshot at 08:13: A comparison of pricing tiers for Z.ai's service, showing the 'Lite' plan at $3/month and the 'Max' plan up to $100/month, contrasting with Claude's $100/month Max plan.](https://ss.rapidrecap.app/screens/82DyXL0ZXI8/00-08-13.jpg)