For your situation, I honestly think the hardware landscape for these big models is kinda frustrating right now. I've spent thousands trying to get smooth inference on the full R1, and unfortunately, it's just not as good as expected on consumer gear because of the sheer weight of the parameters.

• If you want the most bang for your buck, dual NVIDIA GeForce RTX 3090 24GB cards are basically the only way to go. You can find them used for around $750 each. That gives you 48GB VRAM which handles the 67B version at 4-bit quantization pretty easily, but even that setup struggles with the larger R1 models unless you use the distilled ones like the DeepSeek-R1-Distill-Llama-70B.
• I had issues with the Apple Mac Studio M2 Ultra 192GB RAM despite the massive unified memory pool. The memory bandwidth is great on paper, but the actual token-per-second speed for heavy coding tasks was way lower than my multi-GPU rig, which was super disappointing given the $5,000 price tag.
• Right now, the most cost-effective sweet spot is a used workstation with a AMD Ryzen Threadripper 3960X and two or three 3090s. It's loud and eats power like crazy, but it's basically the only way to avoid the massive bottlenecking I saw on single-card setups. Basically, don't expect the full 671B R1 to run locally without an enterprise budget... it's just too heavy for us mortals right now. Anyway, stick to the used 3090s for the 67B and you'll be fine. gl!

Just saw this and honestly dont waste ur cash. I spent way too much on gear before realizing 4-bit quants are the way to go for R1.

• Option A: NVIDIA GeForce RTX 3090 24GB (x2)
• Faster for coding.
• Option B: Apple Mac Mini M4 Pro 64GB
• Better VRAM value. Pro tip: Use LM Studio to check quants first. Learned that the hard way SO many times lol. Peace

I have been obsessing over this for the last few weeks while trying to figure out my own setup. Tbh the market is wild right now and it feels like there is a huge divide between the two main ecosystems.

• NVIDIA is basically the gold standard because of software support, meaning most DeepSeek libraries are optimized for their chips first.
• Apple is really disrupting things with their unified memory architecture which lets you access way more memory for cheaper than top tier enterprise cards. Before you pull the trigger, I have a couple questions to help narrow it down:
• How much are you looking to spend in total?
• Is your priority super fast response times for coding or just being able to run the biggest possible model version regardless of speed? The market for used parts is also shifting really fast, so it depends if you want brand new gear or if you are okay with hunting for deals...

Works great for me

> Should I prioritize multiple 3090s or go for a high-end Mac? What's the most cost-effective hardware setup for smooth inference right now? ^ This. Also, I have been dealing with this exact same dilemma for over a month now and honestly I am still no closer to a real answer. Its so frustrating trying to weigh the massive VRAM capacity of those unified memory systems against the raw speed of a multi-GPU build. Ngl every time I think I have finally made a choice I find a new benchmark that makes me second guess everything... I am basically just stuck in research hell right now.

@Reply #7 - good point! I totally feel your pain on this. Catching up on this thread now and honestly its ridiculous how much of a nightmare it is to just get decent inference without dropping a small fortune. I spent like three weeks straight tweaking configs only to have the hardware feel totally outdated the next month. It drives me crazy how these companies just keep pushing newer, heavier stuff without any regard for those of us actually trying to build local setups that dont require a professional data center budget. It feels like a scam sometimes, paying these premiums just for the privilege of keeping things offline. Anyway, before I dive into my own headaches, how big of a context window are you actually planning on hitting for your code? That usually changes the whole game for me, and I'm curious if you're aiming for full repo analysis or just single file completions.

Just wanted to say thanks for everyone chiming in. Super helpful discussion.

Yo, ngl I'm still learning the ropes with DeepSeek but I've built rigs for years. I think those R1 models are way heavier on VRAM than the 67B ones. IIRC you're gonna need massive memory for smooth inference.

• dual powerful cards for speed
• a Mac for the huge memory pool
• loads of system RAM Honestly, I'd go with the Mac cuz it's usually less of a headache for beginners. gl!

Late to the party but I've been running these big quants for years. Ngl, if you want to run the heavy R1 stuff without the headache of heat and power spikes from multiple consumer cards, check out the NVIDIA RTX A6000 48GB. In my experience, single-card stability wins every time when youre deep in a coding session. That 48GB VRAM pool handles the 67B version perfectly and can squeeze in decent quants of the R1 models without breaking a sweat. If you want a quieter life and need to run the larger quants, look for a refurbished Apple Mac Studio M1 Ultra 128GB RAM. I've tried many setups over the years and for VRAM capacity per dollar, Apples unified memory is basically the easiest route for local inference right now. Just keep in mind NVIDIA is still king if you ever want to fine-tune your own models. But for just running DeepSeek while you work? The Mac is dead silent and just works.