Moonshot Kimi K3 GPU Requirements: Confirmed VRAM & Deployment Guide

Kimi K3's official Hugging Face repository confirms a 1.56TB footprint for the native weights. Since the model shipped within the last day, figures are still settling across the ecosystem, and an official active-parameter count hasn't been published yet, so any specific number you've seen there is a community estimate rather than a confirmed spec.

Moonshot released Kimi K3's open weights on July 27, 2026. Given how recent that is, hardware figures for the model are still settling across the ecosystem. Here's what's confirmed as of today, sourced directly from Moonshot's own repository.

Kimi K3 in One Paragraph: What Shipped Yesterday

Kimi K3 is Moonshot AI's successor to the Kimi K2 family, a 2.8-trillion-parameter sparse mixture-of-experts model with native vision support and a 1-million-token context window. It's positioned as a direct open-weight challenger to closed frontier models, with Moonshot's own launch materials showing K3 leading or tied for first on 8 of 33 public benchmark rows, including strong results on agentic coding tasks. The model routes through 896 total experts, activating 16 per token, and ships under Moonshot's own "Kimi K3 License," described as MIT-like with commercial gates.

Kimi K3 VRAM Requirements: The Confirmed Number

Because K3 shipped so recently, it's worth confirming exactly which figure you're working from. A 594GB estimate for K3's weights has circulated in some early coverage, calculated by unpacking Moonshot's native MXFP4 training format into BF16. That was a reasonable way to estimate the footprint before the weights were public, but the confirmed number now comes directly from the released repository.

1.56TB of native weights puts K3 firmly outside any single-node conversation. Even an 8-GPU H100 node (640GB combined) covers roughly 41% of that footprint, and that's before KV cache for anything approaching the 1M-token context window is added on top.

The Missing Active-Parameter Number

K3 activates 16 of 896 experts per token, a genuinely extreme sparsity ratio, under 2% of experts fire on any given forward pass. What Moonshot has not published is a specific active-parameter count in billions, unlike some other MoE releases that state this figure directly in their model card.

Regardless of the exact active count, the underlying principle holds the same way it does for every MoE model: active parameters describe compute per token, not memory footprint. All 2.8T total parameters, spread across all 896 experts, have to be addressable in memory, since the router can select any expert on any forward pass. A low active-parameter count makes K3 fast to generate tokens once loaded. It has no bearing on whether the model fits on your hardware in the first place.

Kimi Delta Attention: What Actually Changed From K2

The architecture story matters as much as the parameter count here. K3 introduces Kimi Delta Attention (KDA), a linear-attention mechanism using short convolutions, normalized keys, and gated state updates in place of a standard quadratic attention matrix. Paired with Attention Residuals, KDA is what makes K3's 1M-token context window computationally viable rather than a theoretical maximum nobody can actually use.

Moonshot credits this combination, along with routing changes like Quantile Balancing, for roughly 2.5x better scaling efficiency compared to Kimi K2. That efficiency claim is about compute and training stability, not about reducing the hardware needed to hold the model in memory, which remains governed by the 2.8T total parameter count regardless of how efficiently K3 processes each token once loaded.

Deployment Reality

Given the 1.56TB confirmed footprint, realistic self-hosting starts at cluster scale, not single-node.

Given how new this release is, treat any specific "X GPUs get you Y tokens/second" claim you see right now with caution. Real-world serving benchmarks for K3 are still being established, and the confirmed facts today are the weight size, the architecture, and the sparsity ratio, not yet a settled picture of production throughput per GPU tier.

Decision Framework: Four Questions

Common Mistakes When Sizing Kimi K3

Running Kimi K3 on packet.ai

Given K3's confirmed 1.56TB footprint, a self-hosted deployment is a cluster conversation from the start. packet.ai's GPU Clusters are InfiniBand-connected multi-node fabric built for exactly this scale of model, rather than trying to force a 2.8T-parameter checkpoint onto single-node hardware it was never going to fit on.

If you'd rather evaluate K3 without standing up infrastructure yourself, packet.ai's Token Factory, a managed, OpenAI-compatible inference API, is worth joining the waitlist for. Given how fresh K3 is, using a hosted API to test its actual behavior on your workload before committing to self-hosted infrastructure is a reasonable way to de-risk the decision.

For sizing other current models before deciding on infrastructure, packet.ai's general VRAM requirements guide covers the same kind of sizing math across the broader model landscape.

Frequently asked questions

Last reviewed: July 28, 2026, one day after Kimi K3's open weight release. Figures reflect Moonshot's official Hugging Face repository at time of writing. Given how recent this release is, some details, particularly serving benchmarks and community quantization support, are likely to evolve; check current sources before finalizing a production deployment.