The Chinese AI startup Moonshot AI has released a new open model

The Alibaba-supported Chinese AI startup Moonshot AI has released a new open-source model called Kimi K2. It is claimed to excel in advanced knowledge, mathematics, programming, and agency tasks.

🚀 Hello, Kimi K2! Open-Source Agentic Model! 🔹 1T total / 32B active MoE model 🔹 SOTA on SWE Bench Verified, Tau2 & AceBench among open models 🔹Strong in coding and agentic tasks 🐤 Multimodal & thought-mode not supported for now

С Kimi K2, продвинутая агентская интеллект… pic.twitter.com/PlRQNrg9JL

— Kimi.ai (@Kimi_Moonshot) July 11, 2025

In the announcement, the startup stated that its neural network surpassed Claude Opus 4 on two benchmarks and showed better results than GPT-4.1 from OpenAI on a number of metrics. Additionally, the model has a lower token processing cost:

• $0.15 per 1 million input tokens;
• $2.5 for 1 million weekends.

For comparison, Claude Opus 4 charges $15 for 1 million input tokens and $75 for 1 million output tokens — 100 and 30 times more, respectively. In the case of OpenAI's GPT-4.1, the figures are $2 and $8, respectively.

Kimi K2 is developed based on the Mixture-of-Experts architecture and includes 1 trillion parameters, of which 32 billion are active — representing specialized computational blocks used for specific tasks.

Kimi K2 is available for free through the Kimi app and web interface. The company has opened the source code for two versions:

• Kimi-K2-Base — the base model designed for researchers and developers who need full control over configuration and customization;
• Kimi-K2-Instruct — a fine-tuned version optimized for general use, including chatbots and agent AI scenarios.

The first reviews of Kimi K2 in English and Chinese social networks are mostly positive, although some point out the common issue of hallucinations in the AI segment.

Kimi K2 is so good at tool calling and agentic loops, can call multiple tools in parallel and reliably, and knows "when to stop", which is another important property.

It's the first model I feel comfortable using in production since Claude 3.5 Sonnet. pic.twitter.com/TcEkPlBMuk

— Pietro Schirano (@skirano) July 13, 2025

"This is the first model after Claude 3.5 Sonnet that I feel comfortable with in production," wrote Pietro Shirano, founder of the startup MagicPath, on X.

Open AI — trend

The new model reflects a broader trend in the industry — a shift towards open source, which allows both startups and large technology companies to improve efficiency and accelerate the deployment of their AI products. This approach involves providing public access to the program's source code, enabling third-party developers to:

• modify and refine the architecture of the model;
• correct errors;
• scale the functionality;
• implement custom improvements and adaptations for specific tasks.

Among the renowned Chinese startups, open source is offered by DeepSeek and Tencent.

As for American competitors, Meta and partially Google have gone this way.

At the same time, OpenAI CEO Sam Altman announced the postponement of the launch of its own open AI model for an indefinite period. He cited the need for additional testing in the area of security.

we planned to launch our open-weight model next week.

we are delaying it; we need time to run additional safety tests and review high-risk areas. we are not yet sure how long it will take us.

while we trust the community will build great things with this model, once weights are…

— Sam Altman (@sama) July 12, 2025

Recall that in March 2024, Elon Musk announced the release of the Grok source code.