How to Run DeepSeek V3.2 Locally for Private Coding
DeepSeek V3.2’s open weights let you run a capable coding model entirely on your own hardware, so proprietary code never leaves your network — at the cost of real GPU requirements and setup work. Here is what local deployment actually takes.
Why run DeepSeek V3.2 locally instead of the API?
Three reasons. Privacy: if your code or data cannot leave your infrastructure, a hosted API is a non-starter. Cost predictability: a fixed GPU replaces a variable per-token bill. Control: you pin the exact model version and are not affected when a vendor changes or retires an endpoint.
The hosted DeepSeek V3.2 API is cheaper and simpler for most teams. Local is for those with a hard constraint.
What hardware does it need?
This is the reality check. V3.2 is a mixture-of-experts model with a very large total parameter count, so full-precision inference needs serious datacenter GPUs and lots of VRAM. Most teams run a quantised version instead.
- Full weights: multi-GPU datacenter setup, high cost.
- Quantised (4-bit and similar): fits on fewer, smaller GPUs, with some quality loss.
- Distilled smaller variants: run on a single high-end consumer or workstation GPU, lower quality but usable.
Start with a quantised build to test feasibility before buying hardware. Speed depends heavily on VRAM bandwidth and batching.
How do you set it up?
- Download the weights from the official repository that matches the version you want.
- Serve with vLLM or an equivalent inference server, which handles batching and memory paging.
- Expose an OpenAI-compatible endpoint so your existing tooling connects with only a base-URL change.
- Point your editor or agent at that endpoint and run a benchmark of real tasks.
Local vs hosted DeepSeek V3.2
| Factor | Local | Hosted API |
|---|---|---|
| Privacy | Full control | Vendor policy applies |
| Upfront cost | High GPU spend | None |
| Per-request cost | Near zero | Per token |
| Quality | Depends on quantisation | Full model |
| Maintenance | Yours | None |
What are the practical pitfalls?
Quantisation degrades code quality more than people expect — subtle bugs and worse instruction following show up first. Memory exhaustion under concurrent load is a common surprise. And you own the upgrades: when a new version lands, you repeat the download, quantise and benchmark cycle yourself.
Who should do this?
Teams with strict code-confidentiality rules, regulated industries, and organisations already running GPU infrastructure. If none of those apply, the hosted API is cheaper and better on day one.
Quick Summary
- DeepSeek V3.2 open weights enable fully local, private coding inference.
- Full-precision needs multi-GPU datacenter hardware; quantised builds fit smaller setups.
- Serve with vLLM behind an OpenAI-compatible endpoint for drop-in tooling support.
- Quantisation trades code quality for hardware savings.
- Local is for privacy and control, not for saving money on a small team.
