Kirha Planner: Single-Pass Tool Execution Planning
Kirha Planner is an 8 billion parameter model, fine-tuned from Qwen3-8B, that specializes in generating comprehensive tool execution plans from natural language queries. Unlike traditional agentic systems that involve iterative LLM calls, this model produces an entire Directed Acyclic Graph (DAG) of tool operations in a single pass, significantly improving efficiency and predictability.
Key Capabilities & Differentiators
- Single-Pass Planning: Generates a complete execution plan (DAG) in one LLM call, eliminating back-and-forth interactions and reducing latency.
- Cost Predictability: Plans are known upfront, allowing for accurate cost estimation before execution.
- Parallel Execution: Supports concurrent execution of independent steps and automatic chaining of dependent steps.
- Auditability: Every generated plan includes reasoning traces and a full execution log.
- Output Schema Composition: Leverages
outputSchema (part of the official MCP specification) to understand tool outputs and chain tools by matching output fields to input parameters, avoiding exploratory calls. - Zero-Shot Generalization: Trained using QLoRA (4-bit NF4) on the kirha/planner-dataset, the model learns planning methodology rather than specific APIs, enabling it to generalize to any tool catalog at inference time.
When to Use This Model
This model is ideal for developers building agentic systems where:
- Predictable costs and performance are critical.
- Low latency for complex multi-tool operations is required.
- Structured, auditable execution of tool workflows is a priority.
- You need to orchestrate tools with complex dependencies efficiently.
An official TypeScript SDK is available to facilitate parsing, validation, dependency resolution, and parallel execution of the generated plans.