The trading agent
that asks first.
An open 30B tool-calling model for US stocks, fine-tuned from Muse Glimmer. It fetches real data before every answer and waits for your yes before every order.
Every scenario,
turn by turn.
Real transcripts from the eval set. Step through each tool call, tool result and reply exactly as the model sees them.
Ten tools.
Two need your yes.
OpenAI-style function schemas in the robinhood.* namespace, rendered through Glimmer's ATEM template. Wire them to any broker backend you run.
Running locally
in a minute.
One command with Ollama, llama.cpp or LM Studio. Pass the tool schemas from supermuse/tools.py, wire them to your own broker backend, and you have an agent.
- GGUF
Q4_K_M, runs in ~20 GB RAM/VRAM - OpenAI-compatible tool calls
- Apache 2.0, same as the base model
Built on