Hardware · Memory

How much memory does a model need?

Pick an LLM and a context length. We add up the weights at every quantization, the KV cache that grows with context, and runtime overhead, then show the smallest systems and devices that hold it.

Mellum2.1 Thinking (12B-A2.5B) at 8K context

Quantization Weights KV cache + overhead Total Smallest hardware that holds it
Q2_K Low 4.9 GB 0.7 GB 5.6 GB 578 in all
Q3_K_M Low 6.0 GB 0.7 GB 6.7 GB 510 in all
Q4_K_M Good 7.4 GB 0.7 GB 8.1 GB 451 in all
Q5_K_M High 8.6 GB 0.7 GB 9.3 GB 414 in all
Q6_K High 10.0 GB 0.7 GB 10.7 GB 404 in all
Q8_0 Very high 12.9 GB 0.7 GB 13.6 GB 327 in all
F16 Full 24.3 GB 0.7 GB 25.0 GB 192 in all

Hardware is listed when its usable memory is at least 8% above the total. Macs count roughly two-thirds to three-quarters of unified memory as usable. Each link opens the full “can it run” breakdown with speed.

The rule of thumb

1 · WeightsParameters × bytes eachA 70B model at 4 bits per weight is about 70 × 0.6 ≈ 42 GB. Lower quantization shrinks it and costs some quality.
2 · KV cacheGrows with contextEvery token you keep in context stores keys and values for each layer. Doubling context doubles this part, so long documents cost real memory.
3 · OverheadLeave headroomThe runtime, the operating system and other apps need memory too. Aim for a total at least 10% under the memory your GPU or device can give a model.

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.

Was this page helpful?