DeepSeek V3.2 — VRAM requirements & what it costs to run
Bottom line: at q4_k_m with 8k context you need ≈ 425.9 GiB; smallest single accelerator in our catalog: none — multi-GPU/offload.
Quantization table (GiB)
| quant | weights | total @8k | need @8k | smallest fit |
|---|---|---|---|---|
| native | 642.13 | 707.5 | 707.5 | needs multi-GPU / offload |
| bf16 | 1276.57 | 1405.39 | 1405.4 | needs multi-GPU / offload |
| fp8 | 638.29 | 703.27 | 703.3 | needs multi-GPU / offload |
| q8_0 | 678.18 | 747.15 | 747.1 | needs multi-GPU / offload |
| q6_k | 523.4 | 576.89 | 576.9 | needs multi-GPU / offload |
| q5_k_m | 452.39 | 498.78 | 498.8 | needs multi-GPU / offload |
| q4_k_m | 386.16 | 425.94 | 425.9 | needs multi-GPU / offload |
| q3_k_m | 311.96 | 344.31 | 344.3 | needs multi-GPU / offload |
| q2_k | 267.28 | 295.17 | 295.2 | needs multi-GPU / offload |
Dataset: modelvram.com open dataset (CC BY 4.0 / MIT), retrieved 2026-10-20. "total @8k" already includes KV cache + runtime overhead per dataset methodology; we extrapolate KV linearly by context (formula).
KV-cache scaling at q4_k_m (total GiB)
| context | total |
|---|---|
| 4k | 425.6 |
| 8k | 425.9 |
| 16k | 426.5 |
| 32k | 427.7 |
| 128k | 434.9 |
Cost: rent vs buy vs API
Needs ≥425.9 GiB — no single listed rental instance fits (multi-GPU / offload only). modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending.
Cross-check on record for this family: vLLM official recipes, retrieved 2026-10-20: FP8 ≈ 200 GB, NVFP4 ≈ 170 GB; Thunder Compute blog, 2026-08-12: native checkpoint ≈ 175 GB, 3-bit GGUF ≈ 131 GB. Datasets differ by quantisation bookkeeping — see method.
Running DeepSeek V3.2 with Ollama
Typical q4_k_m GGUF download ≈ 386.16 GiB (weights, dataset figure); expect ≥ 425.9 GiB resident at 8k context. Find tags: ollama.com/search (we do not mirror tag lists — sizes on Ollama are their numbers, ours are dataset-sourced).
FAQ
How much VRAM does DeepSeek V3.2 need?
At q4_k_m with an 8k context it needs about 425.9 GiB including KV cache and runtime overhead; the smallest card in our catalog that fits is a multi-GPU setup. Numbers come from the modelvram.com open dataset (CC BY 4.0, retrieved 2026-10-20) and our published formula (https://llmbudgetlab.com/method/).
Can I run DeepSeek V3.2 on a Mac?
Use the calculator: macOS caps GPU-usable memory well below your spec-sheet unified RAM. We count 2/3 of unified (3/4 above 36GB) as a deliberately conservative rule of ours, not an Apple constant — compare required GiB against usable memory, then confirm on your own machine with sysctl iogpu.wired_limit_mb (why).
Is renting a GPU cheaper than buying one for DeepSeek V3.2?
Rental requires multi-GPU — plug both into the calculator's break-even field; buy price and electricity are yours to enter because we refuse to guess market prices.
**DeepSeek V3.2** needs ~425.9 GiB VRAM at q4_k_m (8k context) → minimum multi-GPU. Check & cost-compare: [LLM Budget Lab](https://llmbudgetlab.com/models/deepseek-v3-2/)or grab the SVG badge →
More: builds under $1000 · under $1500 · used GPUs & break-even · RTX 4090 · RTX 5090