12/07/2026 6:26 pm
Topic starter
Notifications
Clear all
DeepSeek Forum
3
Posts
4
Users
0
Reactions
135
Views
0
im honestly so sick of these cloud API costs just skyrocketing whenever I run a few queries for my coding projects. I cant keep justifying the monthly subscription fees anymore. So I was thinking about building a local rig for DeepSeek instead but im totally lost on the hardware requirements. Is a 3090 enough or do I need to hunt down something beefier? Im in the UK and my budget is like 1500 quid but I dont want to waste cash on overkill if I dont have to. Ive been reading conflicting stuff online and its just stressing me out, like do I need tons of VRAM or is the CPU speed the main bottleneck... does anyone have a setup that actually runs this smoothly?
3 Answers
12
12/07/2026 6:40 pm
In my experience, the VRAM is genuinely everything here. You can forget about the CPU being the bottleneck; once the model is loaded into that precious VRAM, your GPU does the heavy lifting. I have used a few different cards over the years, and if you can snag a used NVIDIA GeForce RTX 3090 24GB GDDR6X, you are golden. Just make sure your PSU is up to the task because those cards spike like crazy. Honestly, check out the Ollama GitHub page or the LM Studio docs if you haven't yet, they make running these locally super easy. If you end up needing more headroom later, you could always toss in a second card, but for now, that 24GB buffer is the sweet spot for most coding models. You really dont need to overthink the CPU side, just get something decent that wont throttle the PCIe lanes.
11
12/07/2026 6:26 pm
Honestly, 1500 quid is tight for a top-tier build but a single NVIDIA GeForce RTX 3090 24GB GDDR6X is basically the gold standard for starting out. I tried running it on my old setup and unfortunately had issues with constant crashes, so dont skimp on the VRAM.
- Prioritize that 24GB VRAM
- Get at least 64GB RAM
- Dont stress about the CPU as much It works pretty well for smaller models tho.
1
12/07/2026 7:10 pm
Can confirm this works. Did the same thing on mine and its been solid ever since.
