Can a Ryzen AI Max+ 395 128GB run Llama 3.2 1B Instruct?
Yes
Yes. At Q8_0 it needs 2.2 GB of the 96 GB available and generates around 112.7 tokens per second at an 8k context window.
The numbers
Llama 3.2 1B Instruct has 1.24 billion parameters across 16 layers. The Ryzen AI Max+ 395 128GB has 128 GB of unified memory at 256 GB/s, of which about 96 GB is left for a model once the operating system has taken its share.
| Quantisation | Total needed | Fits | Speed |
|---|---|---|---|
| Q8_0 | 2.2 GB | yes | 112.7 t/s |
| Q6_K | 1.92 GB | yes | 139 t/s |
| Q5_K_M | 1.8 GB | yes | 155.3 t/s |
| Q4_K_M | 1.67 GB | yes | 175.7 t/s |
| IQ4_XS | 1.59 GB | yes | 192.7 t/s |
| Q3_K_M | 1.54 GB | yes | 204.3 t/s |
| IQ3_XXS degraded | 1.42 GB | yes | 240.6 t/s |
| Q2_K degraded | 1.46 GB | yes | 226.8 t/s |
| IQ2_XXS degraded | 1.27 GB | yes | 304 t/s |
9 of the 9 levels fit. Levels marked degraded are listed for completeness, not as advice.
How long a conversation it holds
At Q8_0, memory rises with the length of the conversation because the attention cache keeps a key and value for every token. The longest window that still fits on this device is 128k tokens.
| Context | Cache | Total | Fits |
|---|---|---|---|
| 4k | 0.13 GB | 2.02 GB | yes |
| 8k | 0.25 GB | 2.2 GB | yes |
| 16k | 0.5 GB | 2.58 GB | yes |
| 32k | 1 GB | 3.33 GB | yes |
| 64k | 2 GB | 4.83 GB | yes |
| 128k | 4 GB | 7.83 GB | yes |
See how fast it feels
Ryzen AI Max+ 395 128GB running Llama 3.2 1B Instruct at Q8_0
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. The same answer on RTX 4090 at Q8_0 would take about 319 ms.
Keep going
Everything the Ryzen AI Max+ 395 128GB runs, the full breakdown for Llama 3.2 1B Instruct, or check a different pairing in the calculator.