Kimi K3は何が公開されたのか：2.8兆パラメータと約1.56TBの意味Moonshot AI released full weights of Kimi K3, a 2.8 trillion-parameter LLM, on…

Moonshot AI has released the full weights of Kimi K3, a large language model with 2.8 trillion parameters, on Hugging Face. The release matters because it puts one of the largest openly distributed models yet directly into the hands of developers and researchers, while also highlighting a widening gap between what can be downloaded and what can realistically be run.

The headline figure is eye-catching by itself. At 2.8 trillion parameters, Kimi K3 sits well above most previously published open-weight systems in nominal scale. More significant than the number, though, is what Moonshot AI actually made available: the full weights usable for inference, rather than only a technical report, a benchmark table, or a hosted API. As of July 29, 2026, the complete official repository measures approximately 1.56 TB.

The source draws a careful distinction that is easy to overlook. There are two separate capabilities in play. One is being able to download and store the model files locally. The other is being able to run the model at usable speed on hardware you own. These are not the same thing. Because the weights are public, anyone with sufficient disk space and bandwidth can archive the model. But saving 1.56 TB of files is a very different proposition from executing them efficiently, and at present this is not a size that ordinary consumer PCs can comfortably load and serve.

The storage figure itself offers a hint about how the weights are encoded. A naive full-precision representation of 2.8 trillion parameters would occupy several terabytes, so a total of about 1.56 TB suggests the released weights are stored in a reduced-precision format rather than full 16-bit floats. This appears consistent with the recent industry trend of distributing very large models in lower-precision numeric formats to keep download sizes and memory footprints manageable. The exact format and architecture details would need to be confirmed against Moonshot AI's own documentation.

Running a model of this size locally is constrained mainly by memory and compute, not disk. Large models must fit their active parameters into fast memory during inference, which for high-throughput use typically means GPU VRAM. Consumer graphics cards currently offer tens of gigabytes of VRAM at most, far below what a model of this class requires without aggressive compression. Techniques such as quantization, which further reduces the number of bits per parameter, along with offloading layers to system RAM or disk, can make it technically possible to run large models on modest hardware, but usually at the cost of substantially slower generation speeds. That trade-off is why the practical experience of "running Kimi K3 at home" is likely to remain limited to well-equipped servers or multi-GPU setups for the foreseeable future.

For context, the open-weight landscape has expanded quickly, with releases from groups such as DeepSeek, Alibaba's Qwen, and Meta's Llama making capable models freely downloadable, and Hugging Face serving as the de facto distribution hub for weights, tokenizers, and configuration files. Kimi is Moonshot AI's model line, and prior entries in that family established the company as a notable participant in the open-weight movement. A 2.8 trillion-parameter release continues the broader push toward publishing very large models rather than keeping them strictly behind proprietary APIs, which is meaningful for reproducibility, fine-tuning, and independent evaluation.

The value of a release like this therefore depends heavily on who is looking at it. For most individuals, the immediate benefit is the ability to inspect, archive, and study the weights, not to spin up a fast local chatbot. For organizations with data-center-grade infrastructure, or for cloud providers and inference platforms, open weights of this scale open the door to self-hosted deployments, customization, and research that a closed API would not permit. It is also worth noting that a 1.56 TB download imposes real bandwidth and storage costs that many users will want to weigh before committing.

In short, Kimi K3's release makes the model genuinely available in the sense that its full weights can be obtained and stored, while practical inference on typical consumer hardware remains unrealistic for now. Keeping those two facts separate is the key to understanding what was actually published.