Can Mac M5 Ultra 256G run Kolibri-1 78B-A3.5B?
Comfortable: GGUF Q8_0 needs 87.1 GB at 4K, batch 1; headroom 104.9 GB.
Text input, FP16 KV cache, one card. Unified-memory capacity uses the GPU guide’s usable limit; negative headroom may require host offload. Speed source settings may differ from this 4K estimate.
Measured runs
No measured run for this pair. Sizing is an estimate, not a runtime test.
Quant memory at 4K
| Quant | Total GB | Verdict | Weight sizing |
|---|---|---|---|
| GGUF IQ1_M | 20.9 | Comfortable | Verified file |
| GGUF IQ2_XXS | 24.0 | Comfortable | Verified file |
| GGUF IQ2_XS | 25.6 | Comfortable | Verified file |
| GGUF IQ3_XXS | 35.4 | Comfortable | Verified file |
| GGUF Q3_K_M | 39.5 | Comfortable | Verified file |
| GGUF IQ4_XS | 45.8 | Comfortable | Verified file |
| GGUF Q4_K_M | 52.2 | Comfortable | Verified file |
| GGUF Q8_0 | 87.1 | Comfortable | Verified file |
Context capacity · GGUF Q8_0
| Context | Total GB | Verdict |
|---|---|---|
| 4K | 87.1 | Comfortable |
| 16K | 87.4 | Comfortable |
| 32K | 87.7 | Comfortable |
| 128K | 89.8 | Comfortable |
Generation speed
GGUF Q8_0: 339 tok/s · Bandwidth ceiling (not measured)