kai-os/Carnice-V3

Carnice-V3: An Agentic Qwen3.8-27B Fine-tune

kai-os/Carnice-V3 is a 27 billion parameter model built upon the Qwen3.8-27B base, specifically fine-tuned for agentic applications within the Hermes Agent ecosystem. It incorporates a merged rank-64 rsLoRA adapter, resulting in a full BF16 checkpoint. The model is designed to handle complex tool interactions and reasoning, utilizing a specific chat template for serializing OpenAI-style function schemas into an XML function-call envelope.

Key Capabilities & Characteristics

• Agentic Fine-tuning: Optimized for use as an engineering agent, particularly within the Hermes Agent runtime.
• Tool Use: Proficient in generating and interpreting tool calls, with a focus on structured XML output for dispatch.
• Qwen3.8-27B Base: Inherits the foundational capabilities of the Qwen3.8-27B model, including its configuration, tokenizer, and multimodal processor metadata.
• Specific Chat Template: Utilizes an unchanged chat_template.jinja from Qwen3.8-27B, supporting <tool_call>, <function=...> XML, <tool_response>, and reasoning_effort.
• Merged LoRA: Contains a merged rank-64 rsLoRA adapter, applied to 496 language-model modules, with frozen vision and MTP modules.

Important Considerations

• Behavioral Limitations: This release did not pass formal behavioral quality gates, showing weaker long-horizon completion and task-level tool-contract performance compared to the base model in internal diagnostics. It is not recommended for unattended, destructive, high-stakes, or production agents without independent evaluation and strong runtime controls.
• Small Training Corpus: The model was trained on a very small, private corpus of eight trajectories, limiting its broad generalization claims.
• Multimodal Aspects: While the complete multimodal base is included, vision behavior was not post-trained.

When to Use This Model

Carnice-V3 is best suited for developers experimenting with agentic workflows and tool-use within a controlled Hermes Agent environment. It provides a specialized foundation for tasks requiring structured tool interaction and reasoning, provided its known limitations regarding long-horizon reliability and task-level tool contracts are carefully managed and evaluated for the specific use case.