armand0e/Muse-Glimmer-30B-Fable-Distill

Model Overview

armand0e/Muse-Glimmer-30B-Fable-Distill is a 30 billion parameter model derived from meta-models/Muse-Glimmer-30B. It has been fine-tuned using QLoRA on agentic coding traces and chat data distilled from Claude Fable 5, incorporating a multilingual dataset slice. The model maintains the base's architecture, tokenizer, and chat template, including its vision tower capabilities, though only the language model was trained.

Key Capabilities & Features

• Agentic Reasoning: Optimized for complex reasoning, particularly in agentic coding scenarios, by distilling traces from Claude Fable 5.
• Specialized Chat Format: Utilizes a unique "Onyx ATEM" harmony-style chat format for reasoning (<|start|>assistant to=self<|message|>) and tool calls (<atem:function_calls> XML).
• Multilingual Support: Includes a multilingual component, with approximately 10% of its training tokens being non-English.
• Tool Use: Designed to handle tool calls and results, though it requires custom parser plugins for vLLM due to its non-standard format.

Training Details

The model was trained using QLoRA on unsloth/Muse-Glimmer-30B-unsloth-bnb-4bit with a 24,576-token context. The training data comprised 651 examples (3.62M tokens), including Fable chat (multilingual), Claude Code agent traces, and Cursor agent traces. Supervision was carefully scoped to preserve the base model's routing for thinking and answering, with 40% of tokens supervised.

Usage Considerations

• Custom Parsing Required: Standard vLLM parsers (≤ 0.28) will mis-parse its unique chat format; custom parser plugins or serving without parsers is necessary.
• Multilingual Limitations: While multilingual, non-English content constitutes a smaller portion of its training data.