Serving DeepSeek V4 with Triton Inference Server
DeepSeek V4 has 1600 billion parameters. Triton Inference Server targets industrial deployment across several models. This page gives the memory required and the platforms that qualify.
What is the minimum memory?
DeepSeek V4 totals 1600 billion parameters. In NVFP4 its weights take about 920 GB including the runtime margin: 1,600 billion parameters × 0.5 byte (NVFP4) = 800.0 GB of weights; × 1.15 runtime margin = 920.0 GB. In FP8 the footprint doubles, to about 1,840 GB (1,600 billion parameters × 1 byte (FP8) = 1,600.0 GB of weights; × 1.15 runtime margin = 1,840.0 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 Triton Inference Server for?
Triton Inference Server targets industrial deployment across several models. See the Triton Inference Server and DeepSeek V4 fact sheets.
On which platforms?
- DeepSeek V4 on H200 SXM server
- DeepSeek V4 on B200 SXM
- DeepSeek V4 on B300 SXM
- DeepSeek V4 on GB300 NVL72
How much memory for DeepSeek V4 with Triton Inference Server?
About 920 GB in NVFP4 for the weights, excluding the attention cache.
Does Triton Inference Server suit DeepSeek V4?
Triton Inference Server serves DeepSeek V4 as soon as the machine offers at least 920 GB of memory, the weight footprint in NVFP4. 4 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.
- DeepSeek V4: 1,600 billion parameters, 49 billion active, 1,048,576-token context, MIT licence. Hugging Face repository (config.json, API, model card), checked on 2 September 2026. Pro variant, 1.6 trillion announced; the repository holds 1,599 billion elements, published in mixed FP4 and FP8 (865 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.