The world's largest open-source model has arrived! Moonshot AI releases the 2.8 trillion-parameter Kimi K3, targeting Anthropic.

```

Moonshot AI officially launches a new round of “open source war” among Chinese large models.

On Thursday, July 16th, Moonshot AI officially released the new generation open-source base model Kimi K3, with a parameter scale of 2.8 trillion, and simultaneously launched API services and developer documentation. According to Xinhua News Agency, this is currently the world’s largest open-source model by parameter count, marking a new step in China’s AI model development.

Kimi K3 is another trillion-parameter flagship open-source model from Moonshot AI, and also the company’s most powerful flagship model to date. Following DeepSeek, Moonshot AI becomes another leading Chinese AI company to open its most advanced model with open weights to developers around the world.

Prior to the official release, some overseas media had already reported on Kimi K3 as a “prelude,” saying that Moonshot AI internally regards Kimi K3 as a key product that benchmarks Anthropic’s latest flagship model Claude Opus 4.8. Its overall performance is expected to reach or even surpass Opus 4.8’s mainstream benchmarks, potentially further narrowing the performance gap between cutting-edge Chinese and US AI models.

Industry insiders believe that with Chinese AI companies such as DeepSeek, Zhipu, and Moonshot AI launching high-performance open-source models, while US leading companies like OpenAI and Anthropic stick to closed model strategies, global AI competition is shifting from “who has the strongest model” to “who has the largest developer ecosystem.”

Moonshot AI releases its strongest open-source model to date

According to Moonshot AI, Kimi K3 uses a mixture-of-experts (MoE) architecture, with a total of 2.8 trillion (2.8T) parameters, and is the company’s largest and most capable base model yet. It supports a 1 million (1M) token context window and natively has visual understanding capability, focusing on long-cycle software engineering (Agentic Coding), complex reasoning, and knowledge work—high-difficulty tasks.

Unlike the previous consumer-oriented Kimi intelligent assistant, Kimi K3 emphasizes developer ecosystem building. This time, Moonshot AI opens the model weights, provides API and development documentation; developers can download it for local deployment, or directly call the official cloud service, accelerating enterprise and developer AI application-building based on K3.

Official evaluation results show that Kimi K3 achieves competitive scores in mainstream benchmarks like code generation, agent task execution, and mathematical reasoning.

Specifically, on SWE-bench Verified (software engineering), LiveCodeBench (programming competition), Tau2 (agent capability), and AIME (mathematical reasoning) benchmarks, Kimi K3 reaches the leading level among open-source models, and in some cases approaches or matches top international closed-source models.

Xinhua News Agency notes that in evaluations, Kimi K3’s overall intelligence is close to cutting-edge closed-source models globally. According to Moonshot AI representatives, training Kimi K3 involved a self-developed underlying model architecture and established a scientific system of model training methods.

This time, Moonshot AI releases Kimi K3 via open weights rather than just API access. This means developers can not only download and deploy the model themselves, and further develop it, but also fine-tune and customize it to their needs.

Global attention before release: Benchmarking Anthropic’s flagship model

In fact, even before Kimi K3’s official debut, overseas media had already started closely following this model days earlier.

Reports say Kimi K3 is internally positioned as one of China’s largest AI models, expected to outperform Anthropic’s Claude Opus 4.8 in multiple mainstream benchmarks, further narrowing the performance gap with Opus 4.8, making it Moonshot AI’s most important model upgrade yet.

Although Anthropic has not disclosed Opus 4.8’s parameter count, industry estimates put it at 1.5 to 2 trillion parameters.

The Financial Times cited insiders saying that while Kimi K3 likely still cannot surpass Anthropic’s super-advanced Fable (which was suspended for safety reasons), it is enough to challenge the long-held belief that “Chinese models lag behind American ones by 8 to 12 months.”

For US AI labs, the bigger challenge may not be model capability, but the business model.

Because Kimi K3 is released with open weights, developers worldwide can download, deploy, and modify it, meaning it could quickly build a developer ecosystem like DeepSeek, maintaining competitive pressure on OpenAI and Anthropic’s closed model strategies.

AI competition focus: Beyond performance, cost and open ecosystem

In the past year, US AI companies have invested hundreds of billions of dollars in AI infrastructure, continually launching more powerful frontier models.

Meanwhile, their commercialization prices continue to rise.

According to Anthropic’s website, starting September this year, Claude Opus 4.8's prices will increase by about 50%: input tokens to $3 per 1M, output tokens to $15 per 1M.

In contrast, Chinese AI companies are taking a different path.

Chinese AI firms like DeepSeek and Moonshot AI continue to release open weight models, allowing enterprise privatized deployment and with inference costs significantly lower than top US models.

For example, Moonshot AI’s previously released K2.6 model costs roughly one-third the price of Claude Opus 4.8 for API calls.

As more enterprises focus on controlling AI costs, many overseas companies are starting to try Chinese models to replace some US models to reduce inference expenses.

Marc Andreessen, co-founder of the well-known US VC Andreessen Horowitz, previously stated that Zhipu released GLM-5.2, which has become the first Chinese model to rival or even outperform US flagship lab models on many open benchmarks.

Open source competition heats up, Chinese AI companies’ value re-evaluated

Since the start of this year, nearly all Chinese AI companies have shifted to an open-source path.

DeepSeek quickly garnered global developer attention with the R1 series; Zhipu, MiniMax, and others have since launched open weight models; now Moonshot AI officially joins this camp as well.

Meanwhile, US leading AI companies still insist on closed-source routes.

OpenAI’s latest GPT series and Anthropic’s Claude flagship models have not opened model weights, but instead rely on API subscriptions to continually boost commercial revenue.

This mode difference has become a new divide in global AI competition.

The Financial Times notes that more and more Silicon Valley investors and tech CEOs now believe the performance gap between Chinese and US frontier AI models is fast narrowing, and what really decides the future competitive landscape may not be model rankings, but who can build an open ecosystem covering developers worldwide.

The capital market is also re-evaluating the value of Chinese AI companies.

According to Financial Times, Moonshot AI is amid a new funding round with an estimated valuation of $31.5 billion; DeepSeek has also started a new round, valued at about $71 billion. In comparison, Anthropic’s latest valuation post-fundraising is $96.5 billion, and OpenAI’s about $85.2 billion.

With Kimi K3’s official release, competition between leading Chinese AI companies and US frontier labs is now extending from purely model capability to comprehensive contests over open-source ecosystems, developer communities, and business models.

Risk warning and disclaimerThe market has risks; investment requires caution. This article does not constitute personal investment advice and does not take into account individual users’ specific investment goals, financial conditions, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular situation. Invest accordingly at your own risk. ```