JUNTO ORCHESTRATOR

A workbench for a small swarm of models working together: load several, give each one a role, and send the same question to all of them. Everything runs on this device.

Building the swarm

• + Upload Pip Model adds a node: an RWKV-7 checkpoint (.pth, .safetensors, .gguf) or an HTMLNLM .pip / .json. The same file loaded twice is shared in memory.
• Tap a node's name to rename it.
• Role is that node's system prompt, e.g. "You are the frog expert. Answer briefly." With a role, a real RWKV-7 node gets System: You are Frogbert. <role> and answers in one line, chat-style. Without one, it simply continues your text, which is handy for seeing raw model behaviour.
• Test runs one node on its own. Remove takes it out of the swarm.
• Each node shows its architecture, size, vocabulary and status, plus its soul if it came from a .pip file.

Asking the swarm

@ Broadcast Query sends your message to every node at once, so you can compare how differently sized, tuned or role-played models answer the same thing.

Sampling (Unified Inference)

• Temp: higher is more varied, lower is steadier. Top-P keeps only the most likely words.
• Max Tokens caps the answer length.
• Presence and Frequency discourage repetition (0 = off; real RWKV-7 models only).

Developer console (⚙ Dev)

Top-K, decay, context and system-prompt options, chat format, and the exact last prompt each node saw, with which node saw it.

A note on what they say

Small models tend to be charming and benign, and they're far too small to help anyone do harm. But now and then they say something strange or offensive. User discretion advised.