How to Deploy DeepSeek V4 Flash Locally

MoE: all 284B parameters must stay resident in memory, while speed is set by the 13B active per token.

Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 1M limit show a dash. Each tier takes whichever file of that bit width the repository has, preferring unsloth dynamic quants, so the exact file name varies by model.

Fits when weights + KV cache + runtime overhead is at or below usable memory, single card. "Offload" on MoE models means it runs with expert weights in system RAM (only attention layers and KV cache stay in VRAM); that speed assumes 70 GB/s RAM bandwidth. The context column is the largest tier that fits.

Expert weights live in system RAM (needs at least 155 GB); VRAM holds only attention layers, shared experts and the KV cache, at least 7 GB at Q4. Speed is bound by RAM bandwidth (estimated at 70 GB/s) and is far slower than a full-VRAM setup.

Device	VRAMGB	Price	RAM neededGB	Est. t/s
Tesla V100 16GB	16	$659	155	11.6
Radeon RX 9070 XT	16	$884	155	10.8
GeForce RTX 5060 Ti 16GB	16	$924	155	10.8
GeForce RTX 5070	12	$956	155	11.3
Radeon RX 7800 XT	16	$966	155	10.8
Tesla V100 32GB	32	$1,141	155	11.6
Arc Pro B60	24	$1,164	155	10.1
Radeon RX 7900 XT	20	$1,244	155	11.1
Radeon RX 6950 XT	16	$1,404	155	10.7
GeForce RTX 5070 Ti	16	$1,576	155	11.6
Radeon RX 7900 XTX	24	$1,738	155	11.4
GeForce RTX 5080	16	$1,868	155	11.7
Arc Pro B70	32	$2,074	155	10.6
GeForce RTX 4080	16	$2,187	155	11.4
Radeon AI PRO R9700	32	$2,244	155	10.8
GeForce RTX 3090	24	$2,520	155	11.7
Ryzen AI Max+ 395 128GB	128	$3,913	155	9
RTX PRO 4000 Blackwell	24	$3,956	155	11.3
GeForce RTX 4090	24	$4,905	155	11.7
GeForce RTX 5090	32	$7,489	155	12.1
DGX Spark 128GB	128	$8,223	155	9.9
GeForce RTX 2080 Ti 22GB	22	—	155	11.2
GeForce RTX 5090 D V2	24	—	155	11.9
Arc Pro B65	32	—	155	10.6
GeForce RTX 4080 SUPER 32GB	32	—	155	11.4
GeForce RTX 5090 D	32	—	155	12.1
Instinct MI100	32	—	155	11.6
Instinct MI50 32GB	32	—	155	11.4
Radeon PRO W7800	32	—	155	10.7
RTX PRO 4500 Blackwell	32	—	155	11.6
A100 40GB PCIe	40	—	155	12
CMP 170HX 40GB	40	—	155	12
GeForce RTX 4090 48GB	48	—	155	11.7
Radeon PRO W7900	48	—	155	11.2
RTX PRO 5000 Blackwell 48GB	48	—	155	11.9
CMP 170HX 64GB	64	—	155	12
Instinct MI210	64	—	155	11.8
RTX PRO 5000 Blackwell 72GB	72	—	155	11.9
RTX PRO 6000D	84	—	155	12
RTX PRO 6000 Blackwell	96	—	155	12.1
Mac Studio M4 Max 128GB	128	—	155	10.7
Mac Studio M5 Max 128GB	128	—	155	10.9
MacBook Pro M4 Max 128GB	128	—	155	10.7
MacBook Pro M5 Max 128GB	128	—	155	10.9

llama-server -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:Q4_K_M -c 8192

vllm serve deepseek-ai/DeepSeek-V4-Flash-0731

File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology