Home›Models
Model sheets
Every local model, measured
Every model sheet is built from measurements, not from its marketing page: the real size of the published files, the context memory computed from the attention pattern it declares layer by layer, and what it costs in tokens to say the same thing in four languages.
The catalogue Measured
Models with a sheet18
Smallest / largest1.03 / 433.83 GiB
Fit in 8 GB5 of 18
Costliest languageDE ×1.62
Measured on2026-08-25
| Model | Size | Compression shown | German cost | Fits in 8 GB | License |
|---|---|---|---|---|---|
| deepseek-v4-flash | 127.28 GiB | IQ4_XS | ×1.58 | no | mit |
| gemma4-12b | 6.63 GiB | Q4_K_M | ×1.33 | no | apache-2.0 |
| gemma4-26b-a4b | 15.78 GiB | Q4_K_M | ×1.33 | no | apache-2.0 |
| gemma4-31b | 17.07 GiB | Q4_K_M | ×1.33 | no | apache-2.0 |
| gemma4-e2b | 2.89 GiB | Q4_K_M | ×1.33 | yes | apache-2.0 |
| gemma4-e4b | 4.64 GiB | Q4_K_M | ×1.33 | yes | apache-2.0 |
| glm-5.2 | 433.83 GiB | Q4_K_M | ×1.49 | no | mit |
| gpt-oss-120b | 58.46 GiB | Q4_K_M | ×1.29 | no | apache-2.0 |
| gpt-oss-20b | 10.83 GiB | Q4_K_M | ×1.29 | no | apache-2.0 |
| llama4-scout-17b | 60.87 GiB | Q4_K_M | ×1.30 | no | other |
| mistral-small-24b | 13.35 GiB | Q4_K_M | ×1.40 | no | apache-2.0 |
| qwen3-1.7b | 1.03 GiB | Q4_K_M | ×1.62 | yes | apache-2.0 |
| qwen3-4b-2507 | 2.33 GiB | Q4_K_M | ×1.62 | yes | apache-2.0 |
| qwen3-8b | 4.68 GiB | Q4_K_M | ×1.62 | yes | apache-2.0 |
| qwen3-coder-30b | 17.28 GiB | Q4_K_M | ×1.62 | no | apache-2.0 |
| qwen3.6-27b | 15.66 GiB | Q4_K_M | ×1.31 | no | apache-2.0 |
| qwen3.6-35b-a3b | 20.61 GiB | Q4_K_M | ×1.31 | no | apache-2.0 |
| qwen3.8-27b | 15.33 GiB | Q4_K_M | ×1.31 | no | apache-2.0 |
Sizes are the real published files at Q4_K_M, our quality floor. Where the author does not publish it, the row falls back to the nearest compression and says which one — a size at one compression compared against another is not a comparison.
This table does not say what runs on your machine. It says how big each model is. Which of them fit in your memory — with the context your work needs, and in your language — is what the Model Finder answers.
Check it against my machine