Kimi K3: China’s Newest Open-Weight Model Aims to Close Gap With US Leaders
Set Trending Topics as a preferred source on Google.
Chinese AI developer Moonshot AI has released its latest model, Kimi K3 — and it is already integrated into the web app of the AI chatbot Kimi.com. With a massive parameter count of 2.8 trillion and the goal of blurring the line between open-weight and closed models, the company could lead the next big wave of AI innovation from China. Just how powerful the model actually is, however, remains to be seen: neither Moonshot AI itself nor independent testers have published any benchmarks so far.
Pricing is aggressive in any case: $0.30 per million input tokens (cache hit), $3.00 (cache miss), and $15.00 per million output tokens, with a maximum context window of 1.048 million tokens.
Two Versions in the Kimi Web App
Users of the Kimi.com web app can already use Kimi K3 in two variants. The K3 · Max version is designed as a flagship all-rounder for chat and agent tasks. The second variant, K3 Swarm · Max, targets massively parallelized applications: it is intended to handle extensive research, batch processing, and other large-scale tasks in a single pass.
Expectations for the model are enormous. According to reports in the Financial Times, Kimi K3 is said to significantly close the gap with the highly praised closed models of industry giants such as OpenAI and Anthropic. Experts believe it could even match or surpass Anthropic’s Opus 4.8 in terms of performance. Whether these predictions prove accurate, however, remains an open question: since no benchmarks — from the company or from independent sources — have been released so far, the true capabilities of Kimi K3 cannot be reliably assessed at this time. The first independent tests should bring clarity in the coming days and weeks.
Moonshot AI: A Rapidly Growing Player
In the short time since its founding, Moonshot AI has established itself as one of the central players in China’s AI landscape. The lab specializes in developing cutting-edge language models and belongs to a new generation of companies challenging the technological lead of Western providers.
The company’s commercial success mirrors its technological growth: following a funding round in May that valued the company at 20 billion US dollars, a new round is now being sought that is expected to reach a valuation of 31.5 billion US dollars.
The Strategic Significance of Kimi K3
What sets Kimi K3 apart from many competitors is its sheer size and its positioning as an open-weight model. Estimates suggest the model has a parameter count of between 2 and 3 trillion, which would make it the largest open-weight model from China.
The Kimi series has already proven that it can compete in the open-source market. The predecessor model, Kimi K2.6, currently ranks among the top 5 open-weight models. Performance analysis, however, paints a nuanced picture of the competitive landscape: while Kimi K2.6 still lags significantly behind Zhipu AI’s GLM series in some benchmarks, it is already on par with the current models from MiniMax and DeepSeek in terms of performance. Whether Kimi K3 can close this gap will only become clear once the still-pending independent benchmarks are released.
The Trend Toward Open-Weight Technology
Moonshot AI’s push comes at a time when the debate over the value of expensive, closed AI systems is being reignited. Security concerns regarding data extraction by major providers, along with the high costs of accessing models like ChatGPT or Claude, are driving many companies to look for alternatives.
The focus is increasingly shifting toward the use of open-weight models. Companies prefer to use powerful base models — such as those from DeepSeek, Z.ai, or indeed Moonshot — and train them specifically for their own particular tasks. Should Kimi K3 live up to the high expectations in upcoming benchmarks, it could shift market dynamics in favor of the more affordable and flexible Chinese open-weight models. Until then, however, the enthusiasm remains a bet on potential — reliable performance data is still outstanding.
