Moonshot AI has unveiled Kimi K3, a 2.8 trillion-parameter open-weight Large Language Model (LLM). According to the Chinese startup, the model matches or outperforms some of the most advanced proprietary AI systems from OpenAI and Anthropic across several software engineering and coding benchmarks.
Independent evaluators have also placed Kimi K3 among today’s frontier AI systems, reflecting a broader trend: open-weight AI is no longer merely catching up with proprietary models. In several practical workloads, it is beginning to compete with them and, in some cases, outperform them.
Why Kimi K3 matters
Mozilla’s latest State of Open Source AI report found that open-weight models have reached near parity with proprietary AI in coding, instruction-following and general knowledge. Closed models continue to lead in reasoning, long-context retrieval and complex agentic tasks.
According to Reuters, Moonshot designed Kimi K3 for advanced reasoning, long-horizon coding and knowledge-intensive work. Rather than competing only on cost or openness, it targets precisely the areas where proprietary AI has traditionally maintained an advantage.
How large is Kimi K3?
One reason Kimi K3 has attracted attention is its scale.
The model contains 2.8 trillion parameters, making it the largest open-weight AI model announced so far.
Parameters are the numerical values an AI model learns during training that allow it to recognise patterns and generate responses. While a larger parameter count does not automatically result in better performance, it generally enables a model to represent greater complexity when paired with an efficient architecture and training process.
According to Moonshot, Kimi K3 is larger than several competing open-weight models:
-
Kimi K3: 2.8 trillion parameters -
MiniMax’s upcoming frontier model: Around 2.7 trillion parameters -
DeepSeek V4 Pro: Around 1.6 trillion parameters -
Meituan LongCat 2.0: Around 1.6 trillion parameters -
Z.ai GLM-5 series: About 744 billion parameters
Leading proprietary models from OpenAI, Anthropic and Google do not publicly disclose their parameter counts.
Kimi K3 also supports a one-million-token context window, allowing it to process exceptionally large amounts of information in a single prompt. This capability is useful for analysing lengthy documents, maintaining context across large software projects and supporting long-running enterprise workflows.
How Kimi K3 performs in benchmarks
Moonshot compared Kimi K3 with several leading US frontier AI models across software engineering benchmarks.
According to Reuters, the company said K3 performed competitively with Anthropic’s flagship Fable 5 while substantially outperforming OpenAI’s GPT-5.5, GPT-5.6 Sol and Anthropic’s Claude Opus 4.8 in graphics processing unit (GPU) kernel optimisation.
GPU kernel optimisation measures how effectively AI generates software that maximises hardware utilisation while reducing computing latency.
According to the South China Morning Post, Moonshot also said K3 outperformed Fable 5 and GPT-5.6 Sol on Program Bench and SWE Marathon, benchmarks that evaluate an AI model’s ability to complete complex software engineering projects with minimal human intervention.
Independent evaluations broadly support Kimi K3’s position near the frontier, although they do not fully match Moonshot’s rankings.
Arena AI ranked K3 first for front-end web interface development, ahead of Claude Fable 5 in five of six software development categories.
Artificial Analysis assigned Kimi K3 an Intelligence Index score of 57, placing it alongside Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5.
However, Artificial Analysis continues to rank Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol above Kimi in overall intelligence, suggesting the leadership race remains highly competitive.
The significance is therefore not that China has produced the world’s undisputed best AI model. Rather, Chinese companies are now building open-weight models close enough to the frontier that the long-standing performance advantage enjoyed by proprietary AI is steadily shrinking.
Open-weight AI is moving from catching up to competing
Kimi K3 is not an isolated development.
Over the past year, several Chinese AI companies have rapidly narrowed the gap with leading US laboratories. Reuters recently reported that Z.ai’s GLM-5.2 approached frontier proprietary models across coding evaluations.
Mozilla’s report suggests developers are already responding to this shift. By June 2026, the five highest-volume models on OpenRouter by token usage were all open-weight models, including systems from DeepSeek, Xiaomi, Tencent and MiniMax. Developers have already begun trusting open-weight AI for some of the industry’s most demanding workloads even before reasoning performance reaches full parity with proprietary systems.
Kimi K3 suggests those remaining gaps are continuing to narrow.
Is open AI becoming China’s competitive advantage?
For much of the AI race, US companies competed by building the world’s most capable proprietary models. Chinese companies are increasingly competing by making frontier AI more affordable and accessible.
According to the Mozilla report, Chinese open-weight models account for four of the five highest-volume models on OpenRouter by token usage.
By releasing powerful models openly, Chinese companies encourage developers worldwide to build applications around their ecosystems while shifting computing costs onto users’ own infrastructure.
This approach also reduces dependence on cloud services at a time when US semiconductor export controls continue to tighten.
Kimi K3 fits this strategy.
It is larger than previous Chinese open-weight models, approaches frontier proprietary performance across several independent evaluations and will have its model weights released publicly, allowing developers worldwide to download and customise it.
The competition is increasingly shifting from building the single most capable model to building the largest developer ecosystem around it.
What comes next?
Kimi K3 does not make China the leader in artificial intelligence overnight. The strongest proprietary models from Anthropic and OpenAI continue to lead on several overall intelligence benchmarks. However, it does signal that the competitive landscape is changing.
Only a year ago, the debate centred on whether open-weight AI could approach proprietary models.
Mozilla now estimates that average capability gap at just 3.3 percentage points, and Kimi K3 suggests open-weight models are beginning to challenge even the reasoning and software engineering capabilities where closed AI has remained strongest.
If that trend continues, the industry’s biggest competitive advantage may no longer come from building the single most powerful model. Instead, it may come from making frontier AI capable, affordable and accessible enough for millions of developers to build on it.
