vcruz305/Muse-Glimmer-30B-Hermes-Agentic
vcruz305/Muse-Glimmer-30B-Hermes-Agentic is a 30 billion parameter fine-tuned causal Transformer model developed by Victor Cruz, based on Meta's Muse Glimmer 30B. This model is specifically optimized for agentic tool use, teaching it to call one or two tools and then stop, addressing common issues with consecutive tool calls. With a context length of 131,072 tokens, it excels in scenarios requiring precise tool interaction for Hermes Agent and similar OpenAI-style tool scaffolds.
Loading preview...
Overview
vcruz305/Muse-Glimmer-30B-Hermes-Agentic is a specialized fine-tune of Meta's Muse Glimmer 30B, developed by Victor Cruz. This 30 billion parameter model, with a 131,072 token context length, is engineered to improve agentic behavior, particularly for Hermes Agent and other OpenAI-style tool-calling frameworks. It addresses a common issue where base models might excessively call tools, by training it to reliably stop after one or two tool uses.
Key Capabilities
- Optimized Tool Use: Fine-tuned to call 1-2 tools and then stop, preventing consecutive tool-call loops.
- Agentic Performance: Significantly improves performance on simulated and native Hermes agent benchmarks, achieving 16/20 on loop-gate tasks with 0 HIT_CAP and 7/7 on native Hermes tasks.
- Text + Tools: While the base Muse Glimmer includes vision capabilities, this fine-tune freezes vision weights, focusing its 16-bit merge on text and tool interaction.
Good For
- Local Hermes Agent: Ideal for applications using Hermes Agent or similar OpenAI-tools scaffolds where precise tool-calling and stopping behavior is critical.
- Research on Agentic Traces: Useful for researchers studying short, Hermes-shaped agentic traces and tool interaction.
- Downstream Development: Serves as a strong 16-bit checkpoint for further merging, quantization, or supervised fine-tuning for agentic tasks.