hvss/Dispatch-7B
hvss/Dispatch-7B is a 7 billion parameter model, fine-tuned from Mistral 7B Instruct v0.3, designed to act as an agent orchestrator. It specializes in parsing user requests and tool catalogs to generate structured execution plans, routing tasks to the most cost-effective AI worker. This model excels at reducing operational costs by intelligently dispatching tasks, reserving expensive frontier models only for genuinely complex reasoning. It achieves a 97.4% valid plan rate and an estimated 86.7% cost reduction compared to using large models for all tasks.
Loading preview...
Overview
hvss/Dispatch-7B is a 7 billion parameter model, fine-tuned from Mistral 7B Instruct v0.3, that functions as an agent orchestrator. Its primary role is to act as a "tiny AI manager" that directs a team of potentially more expensive AI workers. It processes user requests and a catalog of available tools to output a structured JSON execution plan, determining which tool or model handles each step, in what order, and when escalation to a frontier model is truly necessary.
Key Capabilities
- Cost Optimization: Significantly reduces operational costs by intelligently routing tasks to the cheapest adequate worker, reserving expensive frontier models only for complex reasoning. Achieves an estimated 86.7% cost reduction.
- Structured Planning: Generates valid, parseable, and schema-correct JSON execution plans with a 97.4% success rate, including dependency graphs.
- Tool Agnostic: Trained on randomized tool catalogs, allowing it to plan effectively with any custom set of tools provided.
- Robustness: Maintains high performance even with distractor-heavy or deliberately incomplete tool catalogs.
When to Use This Model
- Agentic Systems: Ideal for developers building agent-based applications where efficient task orchestration and cost management are critical.
- Workflow Automation: Suitable for automating complex workflows by intelligently dispatching sub-tasks to various tools or smaller models.
- Cost-Sensitive Applications: A strong choice for applications where minimizing API costs for large language models is a priority, by using a small model to manage larger, more expensive ones.
Limitations
- Trained exclusively on English data.
- Routing quality is influenced by the teacher model used for distillation.
- The model plans but does not execute; a separate harness is required for tool execution and validation.