The Chinese Kimi K3 model debuted on July 22, 2026, as Moonshot AI's response to the dominance of Western LLMs. The AI Explained video analyzes its capabilities, benchmarks, and potential impact on the artificial intelligence market. Is it really better than Llama 3.1?
Kimi K3 – a new player in the world of open AI models
On July 22, 2026, the Chinese company Moonshot AI introduced Kimi K3 – a new open weights large language model (LLM) that immediately sparked interest among experts. The AI Explained channel published a detailed analysis of this model, comparing it with competitors such as Llama 3.1 or Qwen 2.5. What makes Kimi K3 stand out, and can it truly shift the power balance in the AI market?
Architecture and technical innovations
Kimi K3 introduces several key innovations that may make it an attractive choice for developers and enterprises:
Mixture of Experts (moe) with dynamic routing
The model uses the Mixture of Experts mechanism, where only selected parts of the neural network are activated depending on the task. In the case of Kimi K3 (70B version), we are dealing with 8 experts, of which only 2 are used simultaneously. This solution allows for:
- a reduction in computational costs by approximately 30% compared to dense models (e.g., Llama 3.1 70B),
- better energy efficiency, which is significant in the context of rising AI training costs,
- scalability – the 400B version is currently in beta and available only to selected partners.
Long context (128K tokens)
Kimi K3 handles the processing of long documents (e.g., source code, scientific papers) without losing coherence. Tests conducted by the video author show that the model retains context better than Llama 3.1 70B, especially in tasks requiring long-term memory, such as summarizing books or analyzing complex documents.
Optimizations for the Chinese language
The model has been optimized for Chinese characters, which allows for:
- a reduction in the number of tokens needed to represent text,
- better handling of Chinese dialects (e.g., Cantonese, Shanghainese),
- higher quality of Chinese-English translations compared to the competition (including GPT-4o).
Benchmarks – how does Kimi K3 compare to the competition?
The author of the AI Explained video presented the results of Kimi K3 70B benchmark tests against other models, based on official Moonshot AI benchmarks and independent community tests. Here are the key results:
| Benchmark | Kimi K3 70B | Llama 3.1 70B | Qwen 2.5 72B | GPT-4o (2024) |
|---|---|---|---|---|
| MMLU (general knowledge) | 84.2% | 83.6% | 84.1% | 88.7% |
| GSM8K (math) | 89.1% | 88.2% | 87.5% | 92.0% |
| humaneval (coding) | 82.3% | 79.1% | 80.5% | 87.5% |
| MT-Bench (chat) | 9.1 | 8.9 | 9.0 | 9.3 |
| longbench (long context) | 85.4% | 82.1% | 83.0% | 86.2% |
Benchmark conclusions:
- Kimi K3 outperforms Llama 3.1 70B in most tests, especially in tasks requiring long context and coding.
- In Chinese, it is the best open weights model, better even than Qwen 2.5 72B.
- In English, it slightly trails GPT-4o but is comparable to Llama 3.1.
- The model's weakness is benchmarks requiring creativity (e.g., creative writing).
Performance in practice – live tests
The video author conducted several practical tests to evaluate how Kimi K3 performs in real-world applications:
Code generation
Kimi K3 handled programming tasks, such as implementing algorithms in Python. The code quality was better than in the case of Llama 3.1, although it did not match GPT-4o (there were optimization errors).
Long context understanding
The model was tested on summarizing a 50-page PDF document in Chinese and English. Kimi K3 did not lose the thread, unlike Llama 3.1, which lost coherence after about 30 pages.
Polish language support
Kimi K3 handles basic tasks well (translation, answering questions), but struggles with grammar and idioms. The video author noted that the model incorrectly uses grammatical cases and mistranslates Polish idioms. Support for the Polish language is currently weaker than in Llama 3.1, but Moonshot AI promises improvements in future updates.
Chinese-English translations
Kimi K3 outperforms the competition (including GPT-4o) in translations from Chinese to English, preserving cultural and idiomatic nuances.
License and availability – what you need to know
Kimi K3 is released under the Moonshot Open Model License (MOML), which:
- allows for commercial use, including fine-tuning and distribution,
- prohibits the use of the model for illegal purposes (e.g., generating disinformation, cyberattacks),
- requires attribution (labeling the model as "Kimi K3 by Moonshot AI"),
- restricts use in military applications.
The model is available:
- as weights on the Hugging Face platform,
- via API on the Kimi Chat platform (free limits for developers),
- with fine-tuning tools (e.g., Kimi-Tune).
Limitations:
- The API is not available in some countries (e.g., the USA due to sanctions),
- model weights are available globally but may be subject to local regulations.
Potential impact of Kimi K3 on the AI market
The premiere of Kimi K3 may have several significant consequences for the artificial intelligence market:
Competition for Llama 3.1
Kimi K3 may dominate the open model market in Asia, especially in China and ASEAN countries. In Europe and the USA, its adoption may be limited due to weaker support for other languages (e.g., Polish).
Impact on the open-source ecosystem
Moonshot AI provides fine-tuning tools, which may accelerate the model's adoption by developers. The video author predicts that Kimi K3 will be integrated with popular frameworks (e.g., Ollama, LM Studio) within the next 2-3 months.
Geopolitical implications
Kimi K3 is further proof of China's growing strength in the field of AI, alongside models such as Qwen (Alibaba) or Ernie (Baidu). This may increase pressure on the USA and EU to accelerate the development of their own open models.
Challenges
- Censorship: The model may be censored regarding political topics (e.g., questions about Tiananmen or Taiwan).
- Security: There is a risk that the model weights will be used for malicious purposes (e.g., phishing, deepfakes).
Frequently Asked Questions
Is Kimi K3 better than Llama 3.1?
Yes, but with caveats:
- In benchmarks, Kimi K3 70B outperforms Llama 3.1 70B, especially in long context and coding.
- In practice, Kimi K3 is better in Chinese and English, but Llama 3.1 has better support for other languages (e.g., Polish).
- The choice depends on the use case: Kimi K3 for Chinese/English, Llama 3.1 for multilingualism.
How does Kimi K3 perform in Polish?
The model handles basic tasks well but struggles with grammar and idioms. Support for Polish is currently weaker than in Llama 3.1, but Moonshot AI promises improvements.
Is Kimi K3 fully open-source?
No. Kimi K3 is open weights – the model weights are publicly available, but the MOML license has restrictions (e.g., prohibition of military use). Training data is not available.
What are the alternatives to Kimi K3?
Depending on your needs:
- Llama 3.1: better support for European languages,
- Qwen 2.5: a good choice for Asian markets,
- deepseek v2: efficient and open-source, but smaller (236B),
- GPT-4o: best quality, but closed and expensive.
Summary – is Kimi K3 worth paying attention to?
Kimi K3 is a groundbreaking open weights model from China that introduces several technical innovations (moe, long context, optimizations for Chinese). In benchmarks, it outperforms Llama 3.1 70B, especially in Chinese and tasks requiring long context. Its availability and license are developer-friendly, but with geopolitical limitations.
The model may dominate the Asian AI market and increase competition for Western models, but its global adoption depends on support for other languages. If you are looking for a model for Chinese or English applications, Kimi K3 is worth considering. For users needing support for the Polish language, Llama 3.1 may still be a better choice.
The premiere of Kimi K3 shows that the open AI model market is still evolving, and the competition between China and the West is gaining momentum. It is worth following the further development of this model, especially if Moonshot AI introduces improvements for other languages.
If you are interested in the impact of AI on various areas of life, check out our post on how artificial intelligence is changing science in 2026.
Comments