Run DeepSeek V4.1 Flash locally
DeepSeek V4.1 Flash is 552B parameters (16B active per token — a mixture-of-experts, so it still needs room for all 552B in memory). Here's the VRAM it needs at each quantization, the smallest GPU that fits, and how a self-hosted box compares with the cheapest verified API.
VRAM by quantization
| Quantization | VRAM needed | Fits on | Notes |
|---|---|---|---|
| int4 (Q4) | 331.2 GB | 8× H100 (640 GB) ($13.84/hr) | smallest footprint, minor quality loss |
| int8 (Q8) | 662.4 GB | multi-node (>640 GB) | near-lossless |
| fp16 (full) | 1324.8 GB | multi-node (>640 GB) | reference quality |
VRAM ≈ params × bytes/param × 1.2 (overhead). See the fullmethod and the break-even calculator.
Self-host vs API
Cheapest API for DeepSeek V4.1 Flash is $0.75 / 1M tokens(blended) via DeepSeek. Running it yourself in int4 fits a8× H100 (640 GB) at $13.84/hr — about $10,103.2/month at 24/7. Those two lines cross at roughly 13.5B tokens/month: below that the API wins on cost, above it the dedicated GPU does (assuming you keep it busy). Tune your own volume — and the throughput you measure on your own hardware — in thebreak-even calculator. Cost is only one axis: if your constraint is sovereignty (prompts that can't leave your machine, no provider dependency), the break-even prices that choice, it doesn't make it.
Run it at home
In int4, DeepSeek V4.1 Flash needs 331.2 GB — more than any consumer desktop or Apple machine we track holds, so at home it means a multi-GPU server. Renting cloud GPUs by the hour, or the API, is the practical route. See the machines table for the ceiling.
Get the weights
Quantized builds on the model hubs — links search live, so they track new builds as they appear:
- GGUF builds llama.cpp · Ollama · LM Studio
- AWQ builds GPU serving (vLLM)
- GPTQ builds GPU serving
- MLX builds Apple silicon
- Ollama library one-command local run