Serving Kimi K3 with vLLM
Kimi K3 has 2800 billion parameters. vLLM targets production serving, throughput and concurrency. This page gives the memory required and the platforms that qualify.
What is the minimum memory?
Kimi K3 totals 2800 billion parameters. In NVFP4 its weights take about 1,610 GB including the runtime margin: 2,800 billion parameters × 0.5 byte (NVFP4) = 1,400.0 GB of weights; × 1.15 runtime margin = 1,610.0 GB. In FP8 the footprint doubles, to about 3,220 GB (2,800 billion parameters × 1 byte (FP8) = 2,800.0 GB of weights; × 1.15 runtime margin = 3,220.0 GB.) The attention cache is not calculated on this page. These models’ architectures (compressed latent, linear or sparse attention, Mamba layers) share no per-token formula; it is measured on the machine, with the target context and concurrency [TO BE MEASURED].
What is vLLM for?
vLLM targets production serving, throughput and concurrency. See the vLLM and Kimi K3 fact sheets.
On which platforms?
How much memory for Kimi K3 with vLLM?
About 1,610 GB in NVFP4 for the weights, excluding the attention cache.
Does vLLM suit Kimi K3?
vLLM serves Kimi K3 as soon as the machine offers at least 1,610 GB of memory, the weight footprint in NVFP4. 2 of the 8 platforms in the catalogue meet that bar.
Method and sources
Memory footprints are calculated, not measured: a reading on real hardware may differ depending on the engine and the exact weight format. Every figure below carries its source and the date of its last check.
- Kimi K3: 2,800 billion parameters, 104 billion active, 1,048,576-token context, Kimi K3 licence. Hugging Face repository (config.json, API, model card), checked on 2 September 2026. 2.8 trillion announced, 16 of 896 experts active; the repository holds 2,780 billion elements, published in MXFP4 (1,561 GB).
- Bytes per parameter: NVFP4 0.5 (4 bits per parameter, hence 0.5 byte; 4-bit published weights weigh 0.54 to 0.56 byte per parameter with their scales (DeepSeek V4 Pro 865 GB for 1,599 billion, Kimi K3 1,561 GB for 2,780 billion)); FP8 1 (8 bits per parameter (E4M3), hence 1 byte, block scales not counted). Source: connaissance/faits.yaml, quantifications family, checked on 2026-08-31.
- Runtime margin × 1.15: QDNA operating assumption, not measured: 15% above the weights for activations, buffers and fragmentation.
- “Fits” threshold at 75% of memory: QDNA assumption, not measured, which keeps the remaining quarter for the attention cache and concurrency.