The model Kimi K3 went live on July 16, accompanied by dramatic headlines suggesting China has leapfrogged the United States. With 2.8 trillion parameters, it is the largest model whose weights are slated to be released openly. Some industry watchers are already describing the launch as another “DeepSeek moment.” Yet the most telling admission does not appear in any headline.
Honest positioning from the developer
Buried inside Moonshot’s technical blog is a striking remark: overall performance still trails the strongest proprietary models, specifically Claude Fable 5 and GPT-5.6 Sol. For a company to publicly state that its own product is not the best remains an unusual move, and that candor makes the record-setting scale worth a closer look.
K3 does claim one important victory. In the Frontend Code Arena, where users compare two anonymous outputs side by side and select the preferred result, K3 ranks first—ahead of both Claude and GPT. The catch is that this benchmark measures visual taste, not factual correctness. Reporting that result as “Kimi beats US models” confuses a preference ranking with an accuracy evaluation.
How the benchmarks were really run
Moonshot publishes its measurements, but the footnotes reveal unequal test conditions. Depending on the benchmark, each model ran inside a different agent harness—sometimes KimiCode, sometimes Claude Code, sometimes Codex. Those are distinct starting environments, not a clean race on an identical track. In one coding evaluation, the company acknowledges that Claude Fable 5 encountered fallbacks on 35 percent of tasks, which may have depressed its score. In another, K3 reaches its peak figure only when using a technique that compacts the context at 300,000 tokens; without that context management, the number falls.
An independent assessment from the analysis firm Artificial Analysis underscores how quickly such figures can flip. There, Kimi K3 scores 49 percent on a hallucination metric. That might sound poor, but the scale is inverted: it counts how often the model does not get things wrong. Higher is better, and Claude Fable 5 sits below K3 at 45 percent, with several GPT-5.6 variants trailing further behind. The lesson is simple: before trusting a benchmark headline, check what is being measured and which direction the scale runs.
The real meaning of “open weights”
For regular users the practical impact will be negligible, even once the weights are released. The reason is sheer size. Moonshot trains K3 from the start in a compact number format called MXFP4, which uses roughly four bits per weight. At 2.8 trillion parameters, the model file alone approaches 1.4 terabytes. For comparison, a typical eight-billion-parameter model that runs on a well-equipped laptop occupies about five gigabytes.
Moonshot itself recommends deploying K3 on supernode setups with 64 or more accelerators. A single professional graphics card with 96 gigabytes of memory falls short by more than a factor of ten. “Open weights” therefore does not mean running the model at home; it means researchers, companies, and cloud providers may download, inspect, and host it on their own infrastructure rather than renting access solely through Moonshot’s servers. For users who want AI without the cloud on their own machine, small models purpose-built for that job remain the appropriate choice.
A technical side note: because K3 was trained at low precision from day one, the compact format is the original state, not a model squeezed down after the fact. The usual concern about quality loss during compression does not apply in the same way.
Pricing and a shift in strategy
Billing is handled in tokens, the small text chunks of usually just a few characters. Through the API, Moonshot charges $3 per million input tokens and $15 per million output tokens. Repeated inputs are billed at 30 cents, a substantially lower rate. This pricing means K3 is no longer a bargain option. It costs roughly three times as much as its own predecessor, which charged just under a dollar for input. Compared with Claude Fable 5 at $10 and $50, K3 is clearly cheaper. Claude Sonnet 5, however, still undercuts it at its current introductory price of $2 and $10, matching K3’s pricing only in September. The era of Chinese models competing primarily on price is coming to an end at Moonshot.
There is one more operational detail that is easy to overlook. K3 currently always reasons at its highest effort level. Moonshot plans to add leaner modes later. For now, even a simple question triggers the expensive reasoning work, making answers thorough but costly. Independent measurements clock the model at around 62 output tokens per second, placing it mid-pack for speed.
Privacy and data handling
For anyone trying K3 in the app or on the website, the privacy policy is the section that matters most. Moonshot states plainly that inputs, audio, images, videos, and files are processed to provide and improve the services, explicitly including the training and optimization of its own models. A dedicated toggle to disable training on user content is absent from the policy. It instead points to general data subject rights, exercisable through account settings or by emailing the privacy contact. ChatGPT and Claude, by contrast, offer a switch directly in the settings.
The policy also notes that collected data includes IP address, device identifiers, session and conversation IDs, and—where device settings permit—clipboard data. On storage location, it states only that data may be transferred to servers outside the user’s country of residence. No specific country is named. The provider is Moonshot AI PTE. LTD., based in Singapore, and the policy carries the date July 7, 2025. For harmless test tasks this is acceptable. For company internals, customer data, or private documents, the same caution applies as with any cloud service, only here the provider’s own terms explicitly allow for training.
Trying K3 costs nothing. Two groups stand to benefit most: developers or designers, and anyone who regularly works with very long documents. K3 is strong where code and visuals intersect—in web interfaces, 3D, and animation. The one-million-token context window is a real asset for lengthy texts, and image processing is built in. For the best answer quality or the smoothest experience, the leading US models remain the better choice. Moonshot says so itself. And anyone hoping to put a frontier model on their own computer soon should keep those 1.4 terabytes firmly in mind. The remarkable thing about K3 is less its ranking than the pace of progress: a Chinese company is delivering a model on par with the world’s best and opening the model data. Two years ago, that would have been difficult to imagine.
Sources: www.kimi.com, platform.kimi.ai, www.kimi.com, artificialanalysis.ai