DeepSeek V3 (685B MoE) — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
The complete model. Only 37B parameters activate per token — far cheaper inference than size suggests. Requires enterprise multi-GPU setup (400GB+ VRAM). Typically accessed via API or on-prem clusters. Rivals GPT-4o on most benchmarks.
DeepSeek V3 (685B MoE) needs about 414 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Includes the KV cache at 8K context, so it reads higher than the headline figure.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 226.6 GB | — | Too big | |
| Q3_K_M 3.41 bpw | 293.4 GB | — | Too big | |
| Q4_K_M 4.83 bpw | 414.9 GB | — | Too big | |
| Q5_K_M 5.67 bpw | 486.9 GB | — | Too big | |
| Q6_K 6.56 bpw | 563.1 GB | — | Too big | |
| Q8_0 8.50 bpw | 729.2 GB | — | Too big | |
| F16 16.00 bpw | 1371.4 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. DeepSeek V3 (685B MoE) VRAM calculator →
Get DeepSeek V3 (685B MoE) running
DeepSeek V3 (685B MoE) is not practical on a single consumer GPU: it needs about 414.4 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run DeepSeek V3 (685B MoE)
Install Ollama, then run:
ollama run deepseek-v3Specifications
Curated — A hand-written entry from before this catalogue recorded its sources. The figures are long-standing but their provenance is not on file.
- Parameters
- 685 Billion (37B active)
- Context window
- 128,000
- Architecture
- MoE (256 experts)
- Provider
- DeepSeek
- Licence
- MIT
- Practical deployment
- Multi-GPU server
- Specified at
- Q4_K_M
- System RAM
- 700 GB
- Record updated
- 2024-12-26
Commercial use permitted. No usage restrictions beyond attribution.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Can I run DeepSeek V3 (685B MoE) on my GPU?
DeepSeek V3 (685B MoE) — frequently asked questions
How much VRAM does DeepSeek V3 (685B MoE) need?
About 414 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does DeepSeek V3 (685B MoE) run on an RTX 4090 (24 GB)?
No. DeepSeek V3 (685B MoE) needs about 414 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run DeepSeek V3 (685B MoE) locally?
Install Ollama and run `ollama run deepseek-v3`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.