Chinese Moonshot Unveils Open AI Model Kimi K3 with 2.8 Trillion Parameters

The Chinese company Moonshot AI has introduced Kimi K3 — one of the largest open language models in the world with 2.8 trillion parameters. The developers claim that the new model is already capable of competing with the best commercial AI systems, and in a number of specialized tests, it even surpasses them. The model weights are promised to be released in July 2026.

A few years ago, open large language models significantly lagged behind the closed systems of OpenAI, Anthropic, and Google. Now, this gap is rapidly narrowing. The Chinese company Moonshot AI claims that its new model, Kimi K3, has already reached the level of leading commercial solutions, and in some specialized tests, it has even outperformed them.

Nearly three trillion parameters — but without unnecessary computations

Kimi K3 is built on a Mixture-of-Experts architecture and contains 2.8 trillion parameters. The model initially supports image processing, and the maximum context window size reaches one million tokens.

Moonshot AI refers to Kimi K3 as the first open model belonging to the class of systems with nearly three trillion parameters. However, this impressive scale does not mean that all computational resources are engaged with each request. The architecture includes 896 specialized "experts," but only 16 of them are activated for processing each token. This approach significantly increases output efficiency compared to models that have to use the entire set of parameters for each query.

In everyday tasks, leaders are still ahead

The developers do not hide that, overall, the user experience with Kimi K3 still lags behind models like Claude Fable 5 and GPT-5.6 Sol. They estimate that these systems still provide a higher level of universal performance in daily use scenarios.

Nevertheless, the results of independent tests look much more interesting.

Where complex tasks matter, Kimi K3 takes the lead

In the FrontierSWE benchmark, designed to evaluate the capabilities of models in solving complex software engineering tasks, Kimi K3 received a score of 81.2. In comparison, GPT-5.6 Sol scored 71.3, while Claude Opus 4.8 scored 66.7.

Another measure, SWE Marathon, also favored the new model. Here, Kimi K3 achieved a score of 42, while GPT-5.6 Sol scored 39 and Claude Fable 5 scored 35.

The achievements are not limited to programming. In the BrowseComp test, the model set a new record — 91.2 points. Additionally, Kimi K3 showed the best results among existing frontier-level models in the following tests:

• AutomationBench;
• SpreadsheetBench 2;
• OmniDocBench;
• several visual reasoning tests.

Considering these results, Moonshot AI's new development demonstrates that open models are already capable of competing with the strongest closed systems not only theoretically but also in a number of practically significant areas.

Why it is wise to approach records with caution

High scores in benchmarks do not always signify a similarly noticeable superiority in real-world use. The performance of models in everyday tasks depends on many factors that cannot be fully reflected in standardized tests.

Additional complexity arises from the fact that in some of the mentioned tests, models were run through different agent environments (agent harnesses). This makes direct comparisons of results sometimes incorrect, and the figures require careful interpretation.

Nonetheless, the emergence of Kimi K3 shows that open artificial intelligence models have nearly caught up with the closed developments of leading American companies, and in certain specialized tasks, they are beginning to set a new level of performance.