Can an RTX 4070 run DeepSeek V4 Flash 0731?
Partly
Not entirely. About 2 of its 43 layers fit on the card and the rest run on the CPU, which brings generation down to roughly 2.9 tokens per second.
The numbers
DeepSeek V4 Flash 0731 has 304.18 billion parameters across 43 layers, of which 20.36 billion are read for each token. The GeForce RTX 4070 has 12 GB of VRAM at 504 GB/s, of which about 11.2 GB is left for a model once the operating system has taken its share.
| Quantisation | Total needed | Fits | Speed |
|---|---|---|---|
| Q8_0 | 301.86 GB | no | — |
| Q6_K | 233.16 GB | no | — |
| Q5_K_M | 202.35 GB | no | — |
| Q4_K_M | 171.9 GB | no | — |
| IQ4_XS | 151.36 GB | no | — |
| Q3_K_M | 139.32 GB | no | — |
| IQ3_XXS degraded | 109.22 GB | no | — |
| Q2_K degraded | 119.49 GB | no | — |
| IQ2_XXS degraded | 73.81 GB | no | — |
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 4070, at a quantisation worth using:
- DeepSeek Coder V2 Lite Instruct — 15.71B at Q4_K_M, 9.8 GB, about 232.4 tokens/s
- DeepSeek Coder 7B Instruct V1.5 — 6.91B at Q6_K, 9.88 GB, about 45.8 tokens/s
- DeepSeek Coder 6.7B Instruct — 6.74B at Q6_K, 10 GB, about 45.2 tokens/s
You can also run it as is. Ollama and llama.cpp will put 2 of the 43 layers on the card and the rest on the CPU without being asked. It works, at roughly 2.9 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 171.77 GB against 171.9 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 V4 Flash 0731 at a quantisation worth using is the Apple M3 Ultra 256GB, around $5,599 for the whole machine, running it at Q4_K_M at about 55.8 tokens per second. Check the current price (affiliate link; indicative price reviewed 2026-09-01). Compare every option.
See how fast it feels
GeForce RTX 4070 running DeepSeek V4 Flash 0731, 2 of 43 layers on the GPU
YouWhy does my model use more memory when the conversation gets longer?
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.
Reading your question
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 4070 runs, the full breakdown for DeepSeek V4 Flash 0731, or check a different pairing in the calculator.