It can fit only with CPU/RAM offload under the tested configuration (4-bit quantized, 8K context, Ollama); it is not a clean GPU-resident fit.
It can fit only with CPU/RAM offload under the tested configuration (4-bit quantized, 8K context, Ollama); it is not a clean GPU-resident fit.
The result is a deterministic memory-fit estimate for the exact runtime, quantization and context shown below. It is not a tokens-per-second benchmark.
This answer is generated from STACKELIER's structured hardware, model and runtime data, then evaluated by the deterministic compatibility policy shown below. It is not generated from search-result summaries.
| Estimated model weights | 18 GB |
|---|---|
| Estimated context reserve | 2.16 GB |
| Estimated runtime overhead | 1.80 GB |
| Estimated required memory | 21.96 GB |
| Usable device memory | 21.60 GB |
| Combined usable memory | 63.20 GB |
Use the Stack Finder to combine hardware constraints with privacy, task and professional requirements.
