For the past two years, the dominant narrative has been simple: America built the world’s leading AI models while China tried to close the gap. That story no longer holds.
On 17 July, Chinese startup Moonshot AI released Kimi K3, a 2.8 trillion-parameter model. It is the largest open-weight model ever released, nearly twice the size of DeepSeek’s latest open model. Two days after launch, Moonshot had to suspend new subscriptions. Demand had overwhelmed its servers.
On benchmarks published by Moonshot itself, Kimi K3 beat Anthropic’s Claude Opus 4.8 on 30 of 35 tests. It beat OpenAI’s GPT-5.6 Sol on 19 of 35. Independent rankings from Artificial Analysis place it third overall, behind only Anthropic’s Claude Fable 5 and GPT-5.6 Sol.
Wei Sun, principal AI analyst at Counterpoint Research, told CNN the gap between Chinese and American models has narrowed to three to six months, though it varies significantly by task.
Kimi K3 is not an isolated breakthrough. It is the latest in a series of increasingly competitive Chinese models. DeepSeek’s R1 model shocked the industry in early 2025 by delivering frontier-level performance at a fraction of the reported training cost. Since then, Moonshot’s Kimi line has released version after version, each one closing the distance further.
Bank of America analyst Alex Liu told CNBC that Kimi K3 proves China’s flagship models can still achieve significant advances even under hardware constraints. Alongside DeepSeek and Kimi, models from Alibaba’s Qwen, Zhipu’s GLM and ByteDance’s Doubao now sit near the top of independent rankings such as OpenCompass. No single company appears to dominate China’s AI race anymore. Leadership now shifts depending on the benchmark, task or release cycle.
Unlike OpenAI and Anthropic, which rely heavily on paid APIs, many Chinese AI companies have embraced open-weight releases as a way to expand global adoption despite restrictions on advanced chips. Wider distribution creates a larger developer ecosystem, encourages third-party improvements and increases the likelihood that Chinese models become embedded in products around the world.
What sets these models apart from their American rivals is not just capability. It is openness. Nearly all major Chinese models ship as open weights, usually under permissive licenses. Anyone can download them, retrain them and deploy them commercially, often for a fraction of what it costs to use a closed American model through an API.
An earlier Kimi release was priced at roughly one-seventh the cost of Claude Opus. That pricing, not raw benchmark scores, is what has made Chinese models so attractive to developers and businesses outside China.
This is the real angle. The AI race is not just about which country builds the most capable model. It is about which country controls the infrastructure the rest of the world builds on. If developers in Bangladesh, India and other cost-sensitive markets increasingly adopt free or inexpensive Chinese open-weight models instead of paid American APIs, China does not need to lead every benchmark. It only needs to become the default platform for developers who value affordability and flexibility over cutting-edge performance.
However, there is one important thing to note. Some of Moonshot’s benchmark claims for Kimi K3 came from the company itself, and independent verification takes time. Morningstar analyst Malik Ahmed Khan cautioned that benchmark scores should not be equated with real-world task capability, since open-source developers often optimise specifically to score well on tests.
Analysts have also pointed out that K3’s price advantage over frontier US models is smaller than DeepSeek’s was at launch. It suggests the cost gap between Chinese and American AI may be narrowing from both sides.
Analysts tracking the pace of releases, including researchers cited by Artificial Analysis and Counterpoint Research, estimate that Chinese models could overtake American ones on aggregate performance indexes by late 2027, if current trends continue. Whether that happens depends on chip supply, export controls and how much compute Chinese labs can access going forward. But the assumption that China is playing catch-up no longer matches what is happening in the market.
For a country like Bangladesh, where freelancers, startups and small businesses often choose tools based on cost rather than brand, this shift matters. Cheaper, open Chinese models mean more local developers can build AI products without needing large AI budgets or a US company’s approval.
The real contest may not be over who builds the world’s smartest AI. It may be over whose models become the default tools for everyone else. For developers in countries like Bangladesh, that choice could matter more than who finishes first.
