godqhr1421's picture
Upload README.md
7633c93 verified
|
Raw
History Blame Contribute Delete
1.77 kB
metadata
title: Muse Glimmer 30B Agentic Demo
emoji: 
colorFrom: purple
colorTo: indigo
sdk: static
app_file: index.html
short_description: Safe tool-call demo for Muse Glimmer 30B
models:
  - vcruz305/Muse-Glimmer-30B-Hermes-Agentic

Muse Glimmer 30B Hermes Agentic demo

This Space demonstrates vcruz305/Muse-Glimmer-30B-Hermes-Agentic. The live static interface can call a public, no-auth OpenAI-compatible endpoint serving the model. A self-hosted ZeroGPU Gradio implementation is included in app.py and is ready to activate when GPU hosting is available.

The interface demonstrates the behavior targeted by the fine-tune: answer directly when a tool is unnecessary, or propose one or two Hermes-style tool calls and stop. Tool calls are parsed and displayed, but never executed.

Safety and limitations

  • The Space does not run commands, read files, browse the web, or perform destructive actions.
  • Outputs are generated by research weights and may be incorrect or unsafe.
  • The model is about 30B parameters and requires a 96 GB ZeroGPU allocation in BF16. The account that owns this Space is not currently eligible to create ZeroGPU Spaces, so the static client does not host the weights itself.
  • The model card reports 16/20 on its simulated loop gate and 7/7 on a small native Hermes battery. These results are not a product certification.

API and MCP

The static demo accepts a public, no-auth OpenAI-compatible endpoint and sends requests directly from the visitor's browser. It does not accept or retain API keys. The included Gradio implementation exposes a generate endpoint and launches with MCP support once the Space is switched to Gradio with ZeroGPU hardware.