Can an RTX 5070 run DeepSeek R1?

The numbers

DeepSeek R1 has 684.53 billion parameters across 61 layers, of which 38.57 billion are read for each token. The GeForce RTX 5070 has 12 GB of VRAM at 672 GB/s, of which about 11.2 GB is left for a model once the operating system has taken its share.

Nothing on this ladder fits, including the two-bit levels we do not recommend.

What to do instead

The straightforward answer is a smaller model from the same family. These do fit on a GeForce RTX 5070, at a quantisation worth using:

• DeepSeek Coder V2 Lite Instruct — 15.71B at Q4_K_M, 9.8 GB, about 309.9 tokens/s
• DeepSeek Coder 7B Instruct V1.5 — 6.91B at Q6_K, 9.88 GB, about 61 tokens/s
• DeepSeek Coder 6.7B Instruct — 6.74B at Q6_K, 10 GB, about 60.2 tokens/s

You can also run it as is. Ollama and llama.cpp will put 1 of the 61 layers on the card and the rest on the CPU without being asked. It works, at roughly 1.4 tokens per second, which is fine for a batch job and tiring for a conversation.

Shortening the context window will not rescue this one: at 4k it still wants 385.99 GB against 386.48 GB at 8k, because the weights rather than the cache are what fill the card. The gap here is too large for settings to close.

Or the hardware that does run it

The cheapest device we track that runs DeepSeek R1 at a quantisation worth using is the Apple M3 Ultra 512GB, around $9,499 for the whole machine, running it at IQ4_XS at about 32.6 tokens per second. Check the current price (affiliate link; indicative price reviewed 2026-09-01). Compare every option.

Wait for the first word350 ms

Then writes at1.4 tok/s

Whole answer2 min 4 s

YouWhy does my model use more memory when the conversation gets longer?

Model

Because of the KV cache. Every token you send leaves behind a key and a value vector in each layer of the model, and those stay in memory for as long as the conversation lasts.

The weights are a fixed cost: load a 4-bit 8B model and that is about 4.8 GB, whether you write one word or ten thousand. The cache is the part that grows, and it grows in a straight line with the number of tokens in the window.

How steeply depends on the model's attention design. With grouped-query attention, several query heads share one key-value pair, which cuts the cache by that ratio. Without it, every head keeps its own, and a long context can cost more memory than the weights themselves.

If you are short on memory, the first thing to try is lowering the context window in your runtime, not the quantisation.

Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU.

Keep going

Everything the GeForce RTX 5070 runs, the full breakdown for DeepSeek R1, or check a different pairing in the calculator.