Models
Tiers
A tier is the RAM class a model is meant for. taraengine doctor names the model for the RAM it finds, and (default) marks that model for each tier.
small
suits any machine
16
at least 14 GiB of RAM
32
at least 30 GiB of RAM
48
at least 46 GiB of RAM
The catalog
| Name | Family | Tier | Size | Context | Capabilities | Quantization | KV cache | Source |
|---|---|---|---|---|---|---|---|---|
gemma4:e4b |
gemma4 4B | small (default) | 4.6 GB | 131,072 | tools, thinking, vision | Q4_0 | q8_0 | Hugging Face |
gemma4:12b |
gemma4 12B | 16 (default) | 7.2 GB | 262,144 | tools, thinking, vision | Q4_0 | f16 | Hugging Face |
gemma4:26b |
gemma4 26B-A4B | 32 | 14.6 GB | 262,144 | tools, thinking, vision | Q4_0 | f16 | Hugging Face |
gemma4:31b |
gemma4 31B | 48 | 18.0 GB | 262,144 | tools, thinking, vision | Q4_0 | f16 | Hugging Face |
glm-4.7-flash |
glm-4.7-flash 30B-A3B | 32 (default) | 18.2 GB | 198,000 | tools, thinking | Q4_K | q8_0 | Hugging Face |
qwen3.6:27b |
qwen3.6 27B | 32 | 19.1 GB | 262,144 | tools, thinking, vision | Q4_K_M | q8_0 | Hugging Face |
qwen3.8:27b |
qwen3.8 27B | 32 | 19.0 GB | 262,144 | tools, thinking, vision | Q4_K_M | q8_0 | Hugging Face |
qwen3.6:35b |
qwen3.6 35B-A3B | 48 (default) | 20.4 GB | 262,144 | tools, thinking, vision | Q4_K_M | q8_0 | Hugging Face |
Size is the main file, as taraengine list shows it; the projector and the draft file come on top. KV cache is the type of the model's attention cache: q8_0 holds less memory than f16 at the same window, and an entry keeps f16 where q8_0 cost it more than a tenth of its speed on the RTX 5090 board. Context is the window the model runs at by default.
Each model keeps the license of its authors; read it on the model's Hugging Face page before you use it. Pull one with taraengine pull gemma4:e4b; the documentation says what a pull checks.