The library
Every model that fits, byte-verified
Sizes are summed from the actual files on Hugging Face, speeds are measured on real hardware, and the list of what does NOT fit is published right below the list of what does. That is the whole method. Last verified: 5 August 2026.
What does not fit — and we will not pretend otherwise
MoE 'active parameters' is a compute figure, not a memory figure: all experts must be resident. Anyone selling you these on a desktop box is confusing the two.
- Kimi K3 (2.8T)594GB even at 1-bit — Not on one box, not on four.
- GLM-5.2 (744B)~405GB Int4-Int8Mix — Measured on 4× Spark at 22–42 tok/s — we build that cluster.
- Hunyuan Hy3 (295B)IQ1_M 181.2GB — Two boxes minimum.
- LongCat 2.0 (1.6T)no GGUF exists — Datacentre class.
- Qwen3.8 / Qwen-Max (≈2.4T)— — Not open weights at all — API only.
Questions, answered straight
- How much of the 128GB can models actually use?
- Operators consistently report 119–121GB allocatable after the OS and driver reserve. Our fit calls use ~115GB as the safe working pool, which is why a 104GB quant with KV cache headroom counts as fitting and a 118GB one does not.
- Why do you quote 3-bit quants — doesn't quality collapse?
- DeepSeek V4 Flash is quantisation-aware-trained with its routed experts natively in MXFP4, so repacking is unusually gentle. Unsloth's published figures show ~96% top-1 token agreement at Q4-class sizes. Below Q4 the published record stops — we are measuring that curve ourselves and will publish it.
- What is the single best model to start with?
- For agentic and reasoning work, DeepSeek V4 Flash 0731 at UD-IQ3_XXS — the heaviest thing that runs on one box, verified hands-on at 15.5–16.6 tok/s with 131K context. For coding with a Western licence chain, Laguna S 2.1 at NVFP4. We install both.
- Can I run video and image models alongside a text model?
- Not simultaneously at the large end — the text flagship occupies ~103GB. The practical pattern is switching workloads (models load in tens of seconds) or dedicating a second box to media. The install sets both patterns up.
