Can an Apple M4 Pro 48GB run OTel 2.0 LLM 31B Instruct?
Yes
Yes. At Q4_K_M it needs 19.62 GB of the 36 GB available and generates around 11.2 tokens per second at an 8k context window.
The numbers
OTel 2.0 LLM 31B Instruct has 32.11 billion parameters across 60 layers. The Apple M4 Pro 48GB has 48 GB of unified memory at 273 GB/s, of which about 36 GB is left for a model once the operating system has taken its share.
| Quantisation | Total needed | Fits | Speed |
|---|---|---|---|
| Q8_0 | 33.34 GB | yes | 6.5 t/s |
| Q6_K | 26.09 GB | yes | 8.4 t/s |
| Q5_K_M | 22.84 GB | yes | 9.6 t/s |
| Q4_K_M | 19.62 GB | yes | 11.2 t/s |
| IQ4_XS | 17.45 GB | yes | 12.7 t/s |
| Q3_K_M | 16.18 GB | yes | 13.7 t/s |
| IQ3_XXS degraded | 13.01 GB | yes | 17.2 t/s |
| Q2_K degraded | 14.09 GB | yes | 15.8 t/s |
| IQ2_XXS degraded | 9.27 GB | yes | 24.7 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 Q4_K_M, 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 256k tokens.
| Context | Cache | Total | Fits |
|---|---|---|---|
| 4k | 0.94 GB | 19.46 GB | yes |
| 8k | 0.94 GB | 19.62 GB | yes |
| 16k | 0.94 GB | 19.95 GB | yes |
| 32k | 0.94 GB | 20.61 GB | yes |
| 64k | 0.94 GB | 21.92 GB | yes |
| 128k | 0.94 GB | 24.54 GB | yes |
See how fast it feels
Apple M4 Pro 48GB running OTel 2.0 LLM 31B Instruct at Q4_K_M
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 Q5_K_M would take about 4.9 s.
Keep going
Everything the Apple M4 Pro 48GB runs, the full breakdown for OTel 2.0 LLM 31B Instruct, or check a different pairing in the calculator.