What hardware do you need to run DeepSeek-V4-Flash?

DeepSeek-V4-Flash is a large mixture-of-experts model: about 284B total parameters with roughly 13B active per token, and a context window in the million-token class. The low active-parameter count keeps compute cheap, but the catch is memory. All 284B weights must sit in memory at once, even though only a fraction fire on any given token. So running it locally splits into two questions: do you have the memory, and can your software load the architecture?

The memory math

These figures are DeepSeek's published architecture and community sizing, not our own measurements. At 4-bit the 284B weights land near 160GB; at 8-bit near 285GB, plus the million-token KV cache on top. Community builds report a heavily quantized floor around 81GB and a realistic entry point of roughly 96GB of VRAM or 128GB of unified memory. A single 24GB consumer card cannot hold it.

What actually has the memory

• Apple Silicon is the cleanest consumer path. A 512GB Mac Studio (M3 Ultra) has far more unified memory than needed: a 4-bit build (~160GB) or even 8-bit (~285GB) fits with room for context. Any Mac with 128GB or more clears the entry bar. See the best model for a 512GB Mac Studio for how large-MoE decode speed behaves on that machine.
• Consumer GPUs do not fit it on one card. No single 24 to 32GB card holds 284B weights. Community setups pair two RTX 4090s (or an RTX 6000-class card) with heavy quantization plus CPU and system-RAM offload, which runs but slowly.
• Data center: roughly an H100 80GB at FP8, or multiple H200s for full weights with long context.

Check the fit for your exact card or Mac with the VRAM-fit checker.

The real blocker is software, not memory

As of this writing, no stable release of llama.cpp, Ollama, or MLX can load the V4 architecture; the only local runs depend on experimental community forks. That is why we have not published a signed decode-speed number for V4-Flash: it does not yet load on our benchmark backend. The moment a stable loader ships, we will measure it and post a signed run here, the same way we do for every other model. See the methodology.

What we have measured in the meantime

For DeepSeek models that run today: DeepSeek-R1 8B decodes 134 tok/s on a 24GB RTX 4090 (see which DeepSeek-R1 size fits a 4090), and the DeepSeek-Coder-V2-Lite MoE decodes 268 tok/s on an RTX 5090. Browse every signed run on the cheatsheet.

V4-Flash sizing reflects DeepSeek's published specs and community reports as of July 2026; the linked tok/s figures are our own signed runs. We will add a measured V4-Flash number once a stable backend can load it.