Hermes MoA: stack frontier models into one virtual model for hard turns

The Stack

Steps

1. 1

   Scaffold the MoA preset and validate it

   Write the moa block in config.yaml (the default preset is a fine start: a GPT-5.5 + DeepSeek reference pair with an Opus-4.8 aggregator). CI parses it and checks the preset shape and that the aggregator is a real model, not another moa preset. Running the mixture needs provider keys and is fenced.

   cat > config.yaml <<'YAML'
   moa:
     default_preset: default
     presets:
       default:
         reference_models:
           - provider: openai-codex
             model: gpt-5.5
           - provider: openrouter
             model: deepseek/deepseek-v4-pro
         aggregator:
           provider: openrouter
           model: anthropic/claude-opus-4.8
         reference_temperature: 0.6
         aggregator_temperature: 0.4
         max_tokens: 4096
         enabled: true
   YAML
   ruby -ryaml -e '
   c = YAML.load_file("config.yaml")
   moa = c["moa"] || {}
   abort "BAD: no moa block" if moa.empty?
   dp = moa["default_preset"]
   presets = moa["presets"] || {}
   preset = presets[dp]
   abort "BAD: default_preset names no existing preset" unless preset
   refs = preset["reference_models"]
   abort "BAD: need at least one reference model" unless refs.is_a?(Array) && refs.length >= 1
   refs.each { |r| abort "BAD: a reference is missing provider or model" unless r["provider"] && r["model"] }
   agg = preset["aggregator"] || {}
   abort "BAD: aggregator missing provider or model" unless agg["provider"] && agg["model"]
   abort "BAD: aggregator is itself an moa preset (recursion blocked)" if agg["provider"] == "moa"
   puts "config OK: MoA preset " + dp + " fans out to " + refs.length.to_s + " reference model(s); acting model = " + agg["provider"] + "/" + agg["model"]
   '

2. 2

   Point it only at your hardest turns (the model step, not checked by CI)

   Select the preset with /model default --provider moa, or run one turn with /moa <prompt> (it runs the mixture for that turn, then restores your normal model). Watch your token bill: if the lift is worth roughly double the cost on your tasks, keep it for hard work. The mixture run and its quality are fenced.