13:32
Kimi AI introduces Kimi K2, a “reasoning” model positioned as more capable than GPT-5
12:58
Companies that bet on AI are now rehiring laid-off employees
11:19
A screenless future: what to expect from Apple’s first smart glasses
09:59
GTA VI delayed
13:32
Apple to use Google’s Gemini to power Siri
09:40
XPeng unveils humanoid robot with soft skin and customizable body types
13:32
Kimi AI introduces Kimi K2, a “reasoning” model positioned as more capable than GPT-5
12:58
Companies that bet on AI are now rehiring laid-off employees
11:19
A screenless future: what to expect from Apple’s first smart glasses
09:59
GTA VI delayed
13:32
Apple to use Google’s Gemini to power Siri
09:40
XPeng unveils humanoid robot with soft skin and customizable body types
13:32
Kimi AI introduces Kimi K2, a “reasoning” model positioned as more capable than GPT-5
12:58
Companies that bet on AI are now rehiring laid-off employees
11:19
A screenless future: what to expect from Apple’s first smart glasses
09:59
GTA VI delayed
13:32
Apple to use Google’s Gemini to power Siri
09:40
XPeng unveils humanoid robot with soft skin and customizable body types
13:32
Kimi AI introduces Kimi K2, a “reasoning” model positioned as more capable than GPT-5
12:58
Companies that bet on AI are now rehiring laid-off employees
11:19
A screenless future: what to expect from Apple’s first smart glasses
09:59
GTA VI delayed
13:32
Apple to use Google’s Gemini to power Siri
09:40
XPeng unveils humanoid robot with soft skin and customizable body types
Chinese AI developer Kimi AI has unveiled a new open model called Kimi K2 Thinking — a “reasoning” system designed to handle complex logical tasks and analyze context windows of up to 256,000 tokens.
The model is already integrated into Kimi’s own chatbot, and the company plans to launch an intelligent agent mode based on it in the near future.
According to Kimi AI, K2 Thinking outperforms GPT-5 in OpenAI’s Last Human Exam benchmark and shows stronger results in several coding and information-retrieval tests. The company did not clarify whether the comparison was made against the full version of GPT-5 or a lighter variant.
In the public LMArena leaderboard, the previous Kimi model without the reasoning mode ranked below DeepSeek V3.2, Qwen 3, Grok 4 Fast, and recent models from Google, OpenAI, and Anthropic. The company claims the new version is designed to close that gap.
The model is also available via API:
- Standard mode: $0.60 per 1M input tokens and $2.50 per 1M output tokens
- Turbo mode: $1.15 and $8 respectively
Kimi K2 Thinking has been released openly on Hugging Face. The developers state the model has 1 trillion parameters, with 32 billion active at a time, suggesting a mixture-of-experts style architecture.
Editor’s pick
