Moonshot AI this week introduced Kimi K3, a model with 2.8 trillion total parameters. The source describes it as the largest open-weight model released so far, with full weights expected to be available before July 27.
On storage alone, the barrier is already steep. According to the figures cited in the source, saving all 2.8 trillion parameters in Moonshot AI’s native 4-bit MXFP4 format would require about 1.4TB to 1.5TB of disk space, and that is before inference even starts.
Full weights are expected before July 27
Moonshot AI said the full open weights for Kimi K3 are scheduled to go online before July 27. They will be released under a Modified-MIT license, which allows self-hosting. At the time referenced in the article, neither the complete weights nor community quantized versions such as GGUF had appeared.
MoE design keeps active parameters far below the total count
Kimi K3 uses a mixture-of-experts, or MoE, architecture. In the article’s description, the model contains 896 expert subnetworks. For each token, the system routes computation through 16 selected experts plus shared experts instead of activating the whole model. That is why the number of parameters actually active during operation sits at about 50 billion to 60 billion.
The model also uses Kimi Delta Attention, or KDA, together with Attention Residuals. Moonshot AI said decoding speed can improve by as much as 6.3x under a 1 million-token long-context setting. The context window reaches 1 million tokens, and the model also supports native visual input.
Moonshot AI’s own guidance sets the floor at 64 accelerators
Moonshot AI’s deployment guide states that self-hosting K3 requires at least 64 accelerators. The native MXFP4 weights are about 1.5TB in size, or 3.7 times the roughly 407GB cited for DeepSeek-R1 671B in Q4 quantized form.
If the calculation is limited to GPU memory needed just to hold the weights, the rough hardware counts are:
- H100 (80GB): about 19 GPUs
- H200 (141GB): about 11 GPUs
- B200 (192GB): about 8 GPUs
That still excludes KV cache, which stores conversational context during inference. The source says Moonshot AI had not disclosed the KV cache requirement for a 1 million-token context window.
Hardware cost tops $2.4 million, and power demand reaches data-center levels
On pricing, the article puts an H100 80GB PCIe card at about $25,000 to $33,000, while SXM versions are priced above $35,000 to $40,000. An 8-GPU H100 server is listed at more than $300,000.
Using Moonshot AI’s 64-GPU threshold, that works out to 8 servers with 8 H100 cards each. Hardware procurement alone would come to more than $2.4 million, or over NT$70 million based on the source.
Power demand is another constraint. A single H100 consumes about 700W, so 64 GPUs would draw roughly 45kW just for the accelerators. That is well beyond normal household electrical capacity and squarely in data-center territory.
K2 was still within reach for some individual users
The contrast with the earlier K2 family is sharp. The source says K2 had 1 trillion total parameters and around 32 billion active parameters. A community-built 1.8-bit extreme quantization version for K2 could run at roughly 247GB, making it possible to boot with a 24GB consumer GPU combined with system memory offloading.
K2.5 required more than 380GB of unified memory, while the Q2 version of K2.6 needed about 350GB. In the comparison cited by the article, a 512GB Mac Studio could still run the previous generation with difficulty, but K3 pushes the entry point back to the data-center class.
For ordinary users who want access to Kimi K3, the options described in the source are narrower: use the K2 family instead, or call the official API.