DeepSeek V4 models
3 DeepSeek V4 models from DeepSeek on Hugging Face, from 304.2B to 1650.5B parameters, published by DeepSeek in FP4 + FP8 Mixed, with a 1M-token context. The smallest official model, DeepSeek-V4-Flash-0731, needs about 564 GB of VRAM at its published precision; the cheapest live fit is AMD MI300X at $7.89/hr.
Part of the DeepSeek series · Previous generation: DeepSeek V3
Pick a size
One row per official DeepSeek V4 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| DeepSeek-V4-Flash-0731 | 304.2B | 564 GB | – | – | AMD MI300X × 3 | $7.89/hr | not published for this architecture |
| DeepSeek-V4-Pro-0813 | 1650.5B | 1128 GB | – | – | AMD MI300X × 6 | $15.77/hr | not published for this architecture |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (2)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| DeepSeek-V4-Flash-0731 | 304.2B | FP4 + FP8 Mixed | 564 GB | AMD MI300X × 3 | $7.89/hr |
| DeepSeek-V4-Pro-0813 | 1650.5B | FP4 + FP8 Mixed | 1128 GB | AMD MI300X × 6 | $15.77/hr |
Fine-tunes and community models (1)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| DeepSeek-V4.1-Flash-1017-obliterated | 754.6B | BF16 | 1687 GB | No live fit | – |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.
Run and fine-tune DeepSeek V4
- How to run DeepSeek V4: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune DeepSeek V4: LoRA and QLoRA memory per model, and the license limits on derivatives.
More from DeepSeek
- DeepSeek R1 (14 models)
- DeepSeek V3 (7 models)
- DeepSeek OCR (2 models)
- DeepSeek Coder (6 models)
- DeepSeek V2 (5 models)
- DeepSeek VL (3 models)
- DeepSeek MoE (2 models)