02/06/2026 10:25 am
Topic starter
Notifications
Clear all
DeepSeek Forum
4
Posts
5
Users
0
Reactions
210
Views
0
What's the best graphics card I should buy to run DeepSeek V4 Pro on my own computer? I keep seeing people talk about this model and it looks super cool for my coding projects but honestly I have no clue where to start with the hardware side of things. I'm based in London and I've got about £800 saved up to upgrade my desktop but I don't even know if that's enough money or if I need more.
Do I need a special brand or just a lot of memory? I heard VRAM is important but I don't really get what that means for speed. Sorry if this is a really basic question I just don't want to buy the wrong thing...
4 Answers
12
02/06/2026 10:50 am
Building on the earlier suggestion, you definitely need to prioritize VRAM capacity and memory bandwidth over raw clock speed for DeepSeek V4 Pro! It is all about fitting weights into memory to avoid bottlenecking the bus. To provide a precise recommendation, what quantization level are you aiming for? Some amazing hardware options:
- NVIDIA GeForce RTX 3090 24GB GDDR6X
- NVIDIA GeForce RTX 4070 Ti Super 16GB GDDR6X Honestly, that 24GB buffer is fantastic!
11
02/06/2026 11:30 am
I once tried cramming a big model into a 12GB card and it was a total nightmare... just crashes everywhere. You gotta be careful with that £800 limit tho.
- NVIDIA GeForce RTX 4070 Ti Super 16GB: snappy and brand new, but 16GB might feel cramped for this specific use case.
- NVIDIA GeForce RTX 3090 24GB: hunting used ones is risky, but that 24GB VRAM is basically mandatory for smooth coding.
3
02/06/2026 10:26 am
Unfortunately, £800 is probably gonna fall short for a model of that scale. I had issues with mid-range setups and the results were not as good as expected. Basically, you should just stick with NVIDIA. You cant go wrong with their architecture for local coding projects. Just get any high-memory card from them because VRAM capacity is the only metric that matters for inference.
2
24/07/2026 7:15 am
TL;DR: Hunt for a used NVIDIA GeForce RTX 3090 24GB GDDR6X. It is the only reliable path at your budget point. Honestly, if you want to run DeepSeek V4 Pro without constant OOM errors, you need VRAM volume above all else. I have been running similar weights and the 24GB buffer on the NVIDIA GeForce RTX 3090 24GB GDDR6X works well for larger quants. It is a workhorse, no complaints regarding reliability if you keep your airflow solid. Alternatively, if you prefer buying new, the NVIDIA GeForce RTX 4070 Ti Super 16GB GDDR6X is a snappy, efficient card. It runs cooler and uses less power, but you will definitely feel the 16GB limit when you try to load more complex contexts. For coding, I would personally choose the extra 8GB of VRAM every single time. Save the extra cash for a decent power supply if your current one is weak.
