The largest open-weight AI model is now Kimi K3, a 2.8-trillion-parameter release from China's Moonshot AI. Kimi K3, released by China's Moonshot AI on July 16, 2026, is a Mixture-of-Experts model with 2.8 trillion total parameters, reported to be the largest open-weight model released to date. It scored 57 on the independent Artificial Analysis Intelligence Index, close to the Claude Opus 4.8 tier, and its full weights open on July 27, 2026. ASAP summarizes the announcement from primary reporting.

What "world's largest" actually rests on

Kimi K3 has 2.8 trillion total parameters, the largest of any released open-weight model. Its architecture is a sparse Mixture-of-Experts that activates only 16 of 896 experts per token, so the active parameters used at each step sit around 50 billion. Storing and serving it needs the full 2.8 trillion scale, but the compute per inference step is far smaller, a "large reservoir, narrow channel" design. The context window is 1 million tokens, and Moonshot AI says its Kimi Delta Attention decodes million-token inputs 6.3 times faster than the earlier Kimi K2. Reporting frames the scale as exceeding DeepSeek V4-Pro (1.6 trillion) and Kimi K2 (1 trillion).

Where the benchmarks actually place it

The performance coordinates are not as simple as the scale. On Artificial Analysis's composite index Kimi K3 scored 57, placing it below the top closed models Claude Fable 5 and GPT-5.6 Sol and near Claude Opus 4.8 and GPT-5.5. On the knowledge-reasoning benchmark GPQA-Diamond it reached 93.5. In short, its position is "top of the open-weight field, just below the closed frontier," which is not the same as being the best model in the world. The numbers make clear that largest in scale and largest in performance are two different statements.

The 51% hallucination warning light

One other number from the same evaluation unsettles the scale story. Artificial Analysis measured Kimi K3's hallucination rate at 51%. Next to an impressive composite of 57 sits a signal that roughly half of its outputs may diverge from fact. That contrast is exactly why a deployment decision should not rest on a benchmark total alone. Answering knowledge questions and not inventing facts are different axes, and for work where accuracy is trust, such as customer support or document summarization, the latter matters more than the headline score. K3's two numbers, side by side, show that scale does not automatically reduce hallucination.

The cost paradox of running 2.8 trillion

Open-weight is not the same as cheap, and K3 makes that plain in one figure. Even with freely downloadable weights, self-hosting 2.8 trillion parameters requires many high-end GPUs and matching memory. For most organizations, running this scale in-house is more expensive than calling an API. Moonshot AI's official API price is $15 per million output tokens and $0.30 per million cached input tokens. Here is the paradox: the freedom of "it's open" is fully available only to the few with the infrastructure to carry that weight, while everyone else circles back to paid APIs. K3's spec sheet reveals that the real beneficiaries of open weights are large operators who can absorb the scale and the inference services that resell it.

July 27 is the real test

There is still a ten-day gap between the announcement and the artifact. Moonshot AI has said it will release the full weights on July 27, 2026, and the license is reported to be permissive for commercial use, though the final wording has not been confirmed. The value of open weights is verified at this moment. Whether the published benchmarks are numbers a third party can reproduce, and whether the license truly permits free redistribution and fine-tuning, can only be confirmed once the weights are in hand. Rather than trusting the launch-day scores as given, it is more reasonable to wait for a neutral harness's re-measurement and community fine-tuning results after release.

What Korean organizations should weigh now

For local teams, three practical questions come before the spectacle of scale. First, is there a GPU budget to run 2.8 trillion parameters in-house; if not, the benefit of openness gives way to the same API dependence. Second, how will data-handling and regulatory risk be managed when using a model distributed by a Chinese operator; self-hosting keeps data from leaving the building, which is a real advantage. Third, given a 51% hallucination rate, which tasks should it touch; it can help with verifiable code generation or drafting, but places where factual accuracy is central need review safeguards. K3 becomes a working asset only after these three questions are answered, not because of the phrase "free and world's largest."

What this release leaves behind

Kimi K3 shows that open-weight competition has hit a new peak on the axis of scale. At the same time, a composite of 57, a 51% hallucination rate, and the infrastructure cost of running 2.8 trillion parameters together reveal that "largest" and "most usable" are not the same. As the weight of openness grows, the contest shifts from parameter count to the ability to carry and verify that weight. The independent reproduction after the weights actually open on July 27 will be K3's real report card.

Source: Moonshot AI Kimi K3 release (announced July 16, 2026, full weights slated for July 27, 2026; 2.8 trillion total MoE parameters, 16 of 896 experts active per token, about 50 billion active parameters, 1 million token context, Artificial Analysis Intelligence Index 57, GPQA-Diamond 93.5, 51% hallucination rate, $15 per million output tokens and $0.30 per million cached input tokens), based on the announcement and reporting by TechCrunch, winbuzzer, and The Next Web, summarized by ASAP.