The K3 performance is outstanding. It scored 57 on Artificial Analysis’s Intelligence Index, the most popular benchmark index for AI LLM performance. This I just behind Claude Fable 5 and GPT-5.6 Sol, the top 2 closed models in the world.

Yet Kimi K3 is an open weight model not a closed one which makes this achievement remarkable.

Intelligence Comparison of Frontier Models (Both Closed and Open)

In other benchmark’s like Arena.ai’s Frontend Code leaderboard Kimi K3 came in top with a score of 1,679, beating Anthropic’s Claude Fable 5 (1,631), OpenAI’s GPT-5.6 Sol (1,618).

Kimi K3 can also be considered frontier class. Moonshot AI itself says K3 is “open frontier intelligence”. Though not official if a LLM AI model achieves the score of 50 on the Artificial Analysis index it is considered frontier class. Among the open weight or open models, currently only Moonshot’s Kimi K3 and Z.ai (Zhupu AI)’s GLM-5.2 pass the threshold. NVIDIA’s Nemotron 3 Ultra is the only other non-Chinese model that comes close that the score is 38 so far from reaching frontier class.

There are credible Western open models, but there none of them reach the frontier class like K3 or GLM-5.2. The open frontier models are dominated by the Chinese.

Has the West failed?

This is more of consequence of incentive design than scientific capability. US frontier labs monetize through premium APIs and subscriptions. Releasing their best weights would turn a high-margin product into a replicable input, cannibalizing pricing power when they require frontier intelligence to remain scarce.

Challengers such as US’ Thinking Machines (founded by Mira Murati, OpenAI’s ex CTO), along with upstream suppliers such as NVIDIA, have a reason to open weights, but neither has yet crossed this capability threshold. Europe has declared an appetite for AI sovereignty, but its commercial system has yet to finance an open model of this caliber.

China’s equation is different. Open models are a distribution strategy. They expand developer share, derivatives, token usage, domestic cloud, chip demand, and influence over the emerging AI stack. The laboratory may capture less revenue per token while creating more ecosystem value around the model.

Companies choose open weights over closed models because they need to host models on-premise, retain data control, fine-tune deeply, manage latency, inspect behavior or avoid dependence on one API provider. There will be high demand for these open models in the coming future.

The US therefore faces a deployment gap. Any organization requiring both high capability and model ownership must currently accept a meaningful performance discount by using a lagging Western model like Nemotron, or pass through China’s open-model ecosystem.

Open model’s compound nature

Open model competition has a different production function. A release is not the end product; it becomes the infrastructure for the next base model, fine-tuning efforts, quantizations, adapters and applications.

DeepSeek used its large R1 model to produce a dataset of step-by-step reasoning data, then fine-tuned six smaller open models (four trained from Qwen, two trained from Llama) on that dataset. This helped it create compact versions that inherited much of R1's reasoning ability. Moonshot’s K2.7 Code was explicitly built on K2.6. Hugging Face counted more than 113,000 derivatives within the Qwen family, and over 200,000 models when repositories tagging Qwen were included.

This is something the US and other countries can benchmark and learn from as well. Interestingly, US based Thinking Machines says its open source model Inkling’s design largely follows DeepSeek-V3, while its post-training bootstrap used synthetic data generated by Kimi K2.5. China is no longer just supplying models, it is paving the way for other open models.

In open models, each release enlarges the technical base, generates more derivatives and lowers the cost of the next iteration. China’s AI models are beginning to resemble ecosystem momentum rather than a series of isolated breakthroughs.

Will China continue to dominate?

There are geo-political factors at play. The White House is reportedly again considering restrictions on cutting-edge Chinese models, potentially including liability requirements for US hosts. China, meanwhile, is reported to be considering controls on foreign access to advanced Chinese weights and training data.

This will make business for Chinese models difficult as there is too much competition domestically and if the US intervenes exporting the models outside of China would be catering to a limited market. Open AI and Anthropic are enjoying strong revenue growth but the Chinese AI companies are still trying to find a way to monetize.

Also the restriction on Nvidia GPU and semiconductors are making it even more difficult for the Chinese companies. The models are cutting edge but they lack sufficient computing power so they cannot expand or scale. Kimi 3 had to stop subscriptions because they could not handle the high demand. Without sufficient compute AI companies cannot make money. This is the catch. China has the best software but hardware is crippled.

In the future the Chinese open models may experience competition not from the US but Taiwan, India, UAE, Saudi as each country tries to develop their own AI. Those new entrants will not be able to copy Anthropic or Open AI’s strategy but the Chinese models have shown how to develop ‘good enough’ AI at a very low cost.

For enterprises, open-model strategy will become a supply chain strategy. Teams should secure and validate licensed weights, keep inference layers portable, maintain credible substitutes and quantify the cost of forced migration before a model becomes embedded across agents, fine-tunes and workflows.

For investors, restrictions could protect the pricing power of US closed labs while raising downstream inference costs and slowing AI diffusion. Value would shift toward proprietary API owners and sovereign inference infrastructure. Open-model hosts and dependent application layers would absorb the disruption. China would sacrifice some global distribution, but gain something strategically revealing – confirmation that its models have become important enough to contain.

Kimi K3 is therefore more than another Chinese model closing another benchmark gap. It marks the moment open intelligence itself becomes a geopolitical dependency. The next phase of the AI race will be decided not only by who builds the smartest model, but also by who can become the fast follower. There will be two paths to choose from.

For more analysis on AI, view AI 360 Pulse - Industry Trends, Intelligence and Impact, May 2026
A curated monthly AI newsletter that distills the most important developments in AI, analyzing their impact across key ecosystem players: semiconductors, big tech, cloud players, AI models, regulators, and users. Designed to cut through the noise, it connects the dots to help you see where the industry is heading and why it matters.