Computed with the open FitLLM engine — accurate per-layer KV-cache modeling, not a naive estimate. Updated 2026-06.
| Model weights | 28.3 GB |
| KV cache | 3.3 GB |
| Runtime + macOS | 6.9 GB |
| Total used | 44.5 / 64 GB |
| Free | 19.5 GB |
Max context at 8-bit: ~90K tokens. Unified memory is shared by the OS — FitLLM leaves ~20% headroom.
| Quant | Weights | Fits (KV F16) | Used @32K |
|---|---|---|---|
| 4bit | 16.2 GB | ✅ up to 206K ctx | 30.8 / 64 GB |
| 8bit | 28.3 GB | ✅ up to 90K ctx | 44.5 / 64 GB |
| 16bit | 54.3 GB | ❌ won't fit | 73.6 / 64 GB |
Gemma 4 31b interleaves sliding-window (local) and global attention 5:1. The local layers cap their KV cache at the 1024-token window, and the global layers use a different head shape (head_dim 512 vs 256). A naive "all layers × full context × one head_dim" formula over-counts KV cache by several times.
same Mac Models that fit in 64GB: Qwen 3.6 27B, Qwen 3.6 35B-A3B, Gemma 4 e2b, Gemma 4 e4b, Gemma 4 12b, Gemma 4 26b A4B, Gemma 4 31b.
Open math: fitllm-engine (MIT), from official config.json.
All numbers are computed by the open-source fitllm-engine (MIT) from official model config.json values — reproduce or audit them yourself. Estimates; real usage varies with runtime (llama.cpp / MLX / Ollama), driver and display. Found a mismatch? Report it. · FitLLM home