Can an RTX 4060 run Llama 3.1 8B Instruct?

In practice

It fits, with very little to spare. Llama 3.1 8B Instruct at Q6_K needs 7.14 GB and the GeForce RTX 4060 offers 7.2 GB, a margin of 0.06 GB. That is enough to load it and run at about 29.7 tokens per second, and not enough to be relaxed about.

Three things take the margin away. A longer window is the first: at 16k the cache alone grows by 1.00 GB. Another program holding VRAM is the second, since a browser with hardware acceleration can take a gigabyte. And a runtime that places layers conservatively is the third, which Ollama does. Store the cache at 8 bits and close the other programs before you load it, and go down one quantization if it still will not.

Reading a 4,000-token prompt, roughly a long email or a few pages of text, takes about 3 seconds before the first word appears. That part is limited by compute and not by bandwidth, so it follows a different spec from the writing speed.

The numbers

Llama 3.1 8B Instruct has 8.03 billion parameters across 32 layers. The GeForce RTX 4060 has 8 GB of VRAM at 272 GB/s, of which about 7.2 GB is left for a model once the operating system has taken its share.

8 of the 9 levels fit entirely. Levels marked degraded are listed for completeness, not as advice. Rows marked partial or experts in RAM run with part of the model in system RAM, at the speed shown.

How long a conversation it holds

Memory rises with the length of the conversation because the attention cache keeps a key and a value for every token. At Q6_K, the longest window that still fits on this device is 8k tokens. Setting the window to what you need and no more is the cheapest way to buy back memory.

The settings we would use

llama-server -m <model>.gguf -ngl all -c 8192 -fa on -ctk q8_0 -ctv q8_0

The cache is stored at 8 bits because headroom is under 1 GB. That halves it for a small quality cost.

In Ollama, set OLLAMA_CONTEXT_LENGTH to the window you want, run ollama ps while the model is loaded and check that the processor column reads 100% GPU. If it shows a split on a model that fits, the Ollama guide shows how to fix it. The llama.cpp flags guide explains each option above.

What else an RTX 4060 runs well

For comparison, these are the other popular models that fit this device at a sound quantization, best pairings first.

The same model on nearby hardware

Memory decides whether Llama 3.1 8B Instruct loads and bandwidth decides how quickly it answers. Here is what a step down, a card with the same memory and a step up do with it.

Wait for the first word355 ms

Then writes at29.7 tok/s

Whole answer6.2 s

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

Model

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

Questions people ask

Keep going

Everything the GeForce RTX 4060 runs, the full breakdown for Llama 3.1 8B Instruct, how the numbers are made in the methodology, or a different pairing in the calculator.