DeepSeek V3.2 685B — VRAM, Speed & Local Setup

Written by Jakub Rusinowski · Last updated March 1, 2026

Model library → DeepSeek V3.2 → DeepSeek V3.2 685B

Updated frontier MoE model with 685B total and 37B active parameters. Delivers ~90% of GPT-5.4 performance. AIME 2026: 91.67% in thinking mode. MIT-licensed. API pricing: $0.28/1M input, $0.41/1M output — among the cheapest frontier models. Self-hosting requires ~370 GB at Q4 (multi-GPU cluster). Major improvement over the original DeepSeek V3 in reasoning and coding.

DeepSeek V3.2 685B 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.

Specifications

Parameters685 Billion (37B active)
Context window128,000
ArchitectureMixture-of-Experts
ProviderDeepSeek
LicenceMIT
Specified atQ4_K_M
System RAM768 GB
Record updated2026-03-01

Licence

MIT — commercial use permitted. Commercial use permitted. No usage restrictions beyond attribution.

VRAM and Speed by Quantization

Modelled on a reference NVIDIA RTX 4090 (24 GB), at 8K context. Speed figures are ESTIMATES from the memory-bandwidth roofline described on the methodology page, not benchmarks we ran — rows marked measured come from published or reader-submitted runs. VRAM here includes the KV cache, so it reads higher than the headline figure above, which does not.

QuantWeightsVRAM neededEst. speedFit on 24 GB
Q2_K225.2 GB226.6 GB—Won't fit
Q3_K_M292.0 GB293.4 GB—Won't fit
Q4_K_M413.6 GB414.9 GB—Won't fit
Q5_K_M485.5 GB486.9 GB—Won't fit
Q6_K561.7 GB563.1 GB—Won't fit
Q8_0727.8 GB729.2 GB—Won't fit
F161370.0 GB1371.4 GB—Won't fit

Want the memory numbers alone, at every quantization level and your own context length? Use the DeepSeek V3.2 685B VRAM calculator.

Buy This HardwareApple Mac Studio M3 Ultra — 512 GB VRAM · 60 W board powerDeploy in the Cloud NowRunPod

or compare on Vast.ai

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.

Recommended GPU

The cheapest catalogued GPU that runs DeepSeek V3.2 685B is the Apple M3 Ultra (512 GB).

Affiliate disclosure: Some links on this page are affiliate links — if you buy through them, LLM Configurator may earn a commission at no extra cost to you. As an Amazon Associate, LLM Configurator earns from qualifying purchases.
Apple Mac Studio M3 Ultra
512 GB VRAM · 60 W board power
2026 prices are volatile — check the current listing.
Check price on Amazon

How to Run DeepSeek V3.2 685B

Install Ollama, then run:

ollama run deepseek-v3-2

Weights on Hugging Face: deepseek-ai/DeepSeek-V3-2.

Best for: reasoning, coding, enterprise, cost efficient.

Can I Run DeepSeek V3.2 685B on My GPU?

Other DeepSeek V3.2 Sizes

DeepSeek V3.2 685B — Frequently Asked Questions

How much VRAM does DeepSeek V3.2 685B 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 V3.2 685B run on an RTX 4090 (24 GB)?
No. DeepSeek V3.2 685B 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 V3.2 685B locally?
Install Ollama and run `ollama run deepseek-v3-2`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.
What other sizes does DeepSeek V3.2 come in?
DeepSeek V3.2 671B (406 GB), DeepSeek V3.2 685B (414 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.

← All DeepSeek V3.2 models | VRAM calculator | Check your own hardware