Most “will it fit” guides stop at the file size. We measure the GGUF, then compute the context cache from the model’s attention shape and add a runtime allowance, and run that across 32 consumer cards and Macs. This week that method produced a result worth sharing: Liquid AI’s LFM2.5 8B-A1B carries almost no context cache, and that changes which card it fits far more than its 4.8 GB file does.
Q4_K_M files, f16 KV cache, llama.cpp-style runtime allowance. Memory in GB.
| Model | Weights | Cache at 8K | Total at 8K | Cache at 32K | Total at 32K |
|---|---|---|---|---|---|
| LFM2.5 8B-A1B | 4.8 | 0.09 | 5.9 | 0.38 | 6.9 |
| Ornith 1.5 9B | 5.4 | 0.25 | 6.7 | 1.0 | 8.2 |
| Qwen3 8B | 4.7 | 1.13 | 6.8 | 4.5 | 10.9 |
| Granite 4.2 8B | 5.0 | 1.25 | 7.3 | 5.0 | 11.8 |
| Gemma 4 12B | 6.6 | 0.81 | 8.5 | 2.31 | 10.7 |
| Phi-4 Mini | 2.3 | 1.0 | 4.2 | 4.0 | 8.0 |
The files are within a gigabyte of each other. The cache at 32K spans a factor of thirteen.
The cache follows the attention shape, which we read from each model’s config and keep in our catalogue:
Mixture of experts does not enter this: the 8.5B total parameters are all in the file, the “A1B” (about one billion active per token, the maker’s figure) affects compute, not memory.
Our verdicts use the card’s usable memory (an RTX 3070 has 7.0 GB usable of 8), not the sticker.
8 GB cards (RTX 3060 Ti, 3070, 4060, 5060)
12 GB cards (RTX 3060 12GB, 4070, 5070)
Further up the ladder for LFM2.5 8B-A1B: 8.3 GB at 64K, 11 GB at 128K (the model’s maximum), which an RTX 4070 Ti Super or any 16 GB card runs comfortably.
All 32 profiles we track fit it at 8K, the Mac mini M6 16GB included.
Model page with the measurement and the per-card table: LFM2.5 8B-A1B VRAM requirements: 5.9 GB at 8K, 6.9 GB at 32K — AI Compute Radar We select a model of the week by a fixed rule: the largest seven-day Heat Score gain among models that run comfortably on a 24 GB card and were not picked in the last four weeks. LFM2.5 8B-A1B went from 28 to 56 between the Friday runs of 9 and 16 October, driven by downloads (26,264 to 32,113 in 30 days). The report with the rule, the frozen readings and the rent-or-buy numbers: Pick of the week, Week 42, 2026: LFM2.5 8B-A1B — AI Compute Radar
If you want the weekly numbers in your inbox, the site has a Monday newsletter that opens with the card you name at signup. Corrections to any number above are welcome here; a correction that survives checking updates the data.