Chinese AI model with open source "Pegaz" that cleans the new GRT-5 and Slaude 4.5

Chinese startup Moonshot AI, backed by Alibaba, is breaking into the world of artificial intelligence with its new model, the Kimi K2 Thinking. Released freely under a modified MIT license, this open source system not only comes close to, but even outperforms, the leading paid models of Orenai, Anthroris, and HAi in key tests of reasoning, programming, and autonomous actions.

Developers can access the model through platform.moonshot.ai, kimi.com, and Hugging Face, where the weights and code are published.

Technical specifications and achievements

K2 Thinking is a Mixture-of-Experts architecture with 1 trillion parameters, of which 32 billion are activated during each infection. The model achieved revolutionary results for all sources: 44.9% on Humanity's Last Eham, 60.2% on Russia (more recently 54.9% for GRT-5 and only 24.1% for Сlaude 4.5), 71.3% on СWE-Венсh Verіfіеd and 83.1% on LiveСодеВенсh v6.

These results put it above the GRT-5 in the judgments and bring it closer in the mathematical tasks, beating even the previous leader among open models MiniMax-M2, which achieved 44.0% on BrowserComp and 69.4% on SWE-Vensh less than two weeks ago.

Autonomous capabilities and efficiency

The defining characteristic is the ability to autonomously use 200-300 consecutive instruments without human intervention. The model preserves consistency in complex multi-step tasks.

A demonstration workflow shows how K2 Thinking autonomously gathers news, analyzes information, and generates structured reports. The model supports a 256k token context window and INT4 quantization, doubling the speed without losing accuracy.

Economical accessibility and licensing

The license allows full commercial use with one condition: for over 100 million monthly users or revenues over $20 million, a visible mention of "Kimi K2" is required.

The economic model is very aggressive – $0.15/1M token (sachet hit), $0.60/1M (sachet miss) and $2.50/1M output, ten times cheaper than GRT-5 ($1.25 input/$10 output).

The custom price is only $4.6 million, due to OrenaII's milliapds and even below $5.6 million for DeerSeek V3.

The needle changes. Again

This achievement comes at a time of growing concerns about the financial sustainability of AI giants. The financial director of Orenai has recently sparked controversy with comments about possible state support for $1.4 trillion investments.

At the same time, companies like Airbnb are already publicly admitting that they use Chinese open alternatives like Qwen.

Chip limitations are not stopping Chinese progress – DeerSeek showed effective training for $5.6 million, and Moonshot is already improving on this result.

When Deerseek first suggested that they could achieve "American" results with a fraction of American budgets, the market was frozen and it took a lot of time, threats of technology bans, and marketing gimmicks from the big players to appease investors.

At the end of last week, the market started to pick up again and it started to feel – and it wasn't just a whisper, it was a "Balloon" smoke.

Source: https://money.bg/