Models

Native, browser, and meta-routing model ids. Every entry lists its tier, size, family, and reward weight.

Catalogue

text
gpu-mesh        auto       —      —       Auto-route to any available model
llama-3.2-1b    light      Llama  0.9 GB  Runs anywhere. Reward 0.6
qwen-2.5-1.5b   light      Qwen   1.3 GB  Sharper reasoning. Reward 0.7
gemma-2-2b      light      Gemma  1.6 GB  Concise. Reward 0.8
llama-3.2-3b    standard   Llama  2.0 GB  Default. Reward 1.0
phi-3.5         standard   Phi    2.2 GB  Strong on instruction. Reward 1.0
qwen-2.5-3b     standard   Qwen   2.0 GB  Reward 1.1
hermes-3-8b     heavy      Llama  4.6 GB  Browser-only. Reward 1.6
mistral-7b      heavy      Mistral 4.4 GB Native-only. Reward 1.8

Live availability

GET /v1/models returns each model with nodes_available. A model can be listed but temporarily unavailable if no live node has it cached. The dashboard's chat dropdown shows this count next to every option.

Reward weights

Bigger models earn more per task. Browser-only entries are weighted slightly lower than native to reflect their lower throughput.

Auto model selection

Requesting gpu-mesh defers the model choice to the orchestrator — any provider with anything cached can serve it. Use this in production unless you have a hard requirement on a specific model family. Requesting a specific id pins the request to a provider that has that model.