Cheapest way to run DeepSeek-V3.2
Quick answer
At 4-bit with an 8K-token conversation, DeepSeek-V3.2 needs about 406.0 GB. The cheapest rental big enough is 8× A100 SXM on Lambda at $25.52 an hour.
| Option | What you need | Cost | Notes |
|---|---|---|---|
| Your own GPU | Doesn't fit any single consumer card in our list (largest 32 GB) at 4-bit with 8K context | – | – |
| Mac (unified memory) | Mac Studio, 512 GB at 4-bit (Q4_K_M), needs 406.0 GB | No Mac prices in our data | $0 per token once you own it |
| Rent a GPU | 8× A100 SXM (640 GB total) on Lambda, On-demand instance | $25.52 per hour ($3.19 per GPU) | Pay only while it runs. Price checked 2026-11-09 |
Fit rule: the Q4_K_M file (or the highest precision that fits) plus an 8K-token FP16 KV cache under 90% of memory, same as the calculator. Sizes and sources: DeepSeek-V3.2 VRAM page. Rental prices: GPU prices (checked 2026-11-09). Hourly and per-token prices measure different things, so which is cheaper depends on how much you use it.
Links marked Referral link, Affiliate partner or Amazon affiliate link may earn us credit or a commission at no cost to you. Disclosure.
Precision ladder: what each quant costs
| Precision | Needs at 8K | Smallest consumer GPU | Smallest Mac | Cheapest single-GPU rental | Cheapest rental overall (any GPU count) |
|---|---|---|---|---|---|
| 16-bit (FP16/BF16) | 1342.9 GB | none (max 32 GB) | none in our Mac list | none big enough in our price data | none big enough in our price data |
| 8-bit (Q8_0) | 713.9 GB | none (max 32 GB) | none in our Mac list | none big enough in our price data | 8× B200 SXM6 (1440 GB) on Lambda, $61.52/hr |
| 4-bit (Q4_K_M) | 406.0 GB | none (max 32 GB) | Mac Studio, 512 GB at 4-bit (Q4_K_M) | none big enough in our price data | 8× A100 SXM (640 GB) on Lambda, $25.52/hr |
Every row names its precision, Mac rows included (Macs share unified memory with the system, so the same 90% rule applies). Single-GPU and multi-GPU rentals are listed separately: a multi-GPU rental needs software that splits the model across cards (llama.cpp, vLLM tensor parallel). Prices checked 2026-11-09.
How context length changes the bill (4-bit)
| Context | Needs | Cheapest rental overall |
|---|---|---|
| 8K tokens | 406.0 GB | 8× A100 SXM (640 GB) on Lambda, $25.52/hr |
| 32K tokens | 408.0 GB | 8× A100 SXM (640 GB) on Lambda, $25.52/hr |
| 128K tokens | 415.6 GB | 8× A100 SXM (640 GB) on Lambda, $25.52/hr |
What $100 of rental buys you
We list no GPU purchase prices (they change daily and we cannot check them), so this page does not compare renting with buying. At the cheapest rental that fits (8× A100 SXM (640 GB) on Lambda, $25.52/hr), $100 of rental buys about 4 hours of running DeepSeek-V3.2 at 4-bit with 8K context. Electricity, storage and setup time are not counted.
Fit quirks for this model
- DeepSeek-V3.2 uses Multi-head Latent Attention: instead of full keys and values, engines such as llama.cpp cache one compressed latent per layer. From config.json (61 layers, kv_lora_rank 512, qk_rope_head_dim 64), an FP16 cache takes 61 × (512 + 64) × 2 bytes = 70,272 bytes per token. The small extra key cache used by the sparse-attention indexer (index_head_dim 128) is not included, and engines Plus the DeepSeek Sparse Attention indexer key cache, stored in FP8: 61 layers × 128 (index_head_dim) × 1 byte = 7,808 bytes per token, for a total of 78,080 bytes per token (a bf16 indexer cache would double that part to 15,616).
- Maximum context is 163,840 tokens; longer settings are capped.
Model data checked 2026-11-03; rental prices checked 2026-11-09. For file sizes, sources and the full fit table see the DeepSeek-V3.2 VRAM requirements page; this page is only about cost.
Get fit alerts for your GPU
Join the fit-alerts waitlist
Alerts are not running yet. Join the waitlist and pick your card once; when alerts launch we will email you once to confirm, and only after you confirm will you get the first alerts. Unsubscribe any time.
More: all cheapest-way pages · cheapest GPU rental by VRAM · Ask VRAMNeeded · Scan my GPU