SZLHOLDINGS/SZL-Khipu-1.5B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

SZL-Khipu-1.5B by SZL Holdings is a compact 1.5 billion parameter model, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct, specifically designed for governed agent navigation. This model proposes retrieval plans as JSON, routing over candidate Brain node handles and citing only those whose metadata supports a given query. It excels at generating schema-constrained JSON proposals for retrieval within a controller boundary, abstaining when no relevant handles are found.

Loading preview...

SZL-Khipu-1.5B: Governed Retrieval Navigator

SZL-Khipu-1.5B is a 1.5 billion parameter model, fine-tuned by SZL Holdings from Qwen/Qwen2.5-1.5B-Instruct, optimized for governed agent navigation. Its primary function is to act as a governed retrieval navigator, proposing retrieval plans in JSON format based on a query and a set of candidate Brain node handles. The model processes only handle metadata (IDs, kind, label, note) and never accesses node content, ensuring that content resolution and execution are handled by an external controller.

Key Capabilities

  • JSON Plan Generation: Emits a single JSON plan conforming to the Khipu output schema, indicating whether to NAVIGATE (cite handles) or ABSTAIN (no relevant handles).
  • Handle-based Retrieval: Routes over provided candidate handles, citing only those whose metadata supports the query.
  • Controller Boundary Integration: Designed to operate within a larger system where an external controller validates plans, resolves handles, and applies governance outside the model's weights.
  • Compact and Deployable: Available in GGUF quants (e.g., Q4_K_M) for CPU-only execution via ollama or llama.cpp, making it suitable for resource-constrained environments.

Limitations and Considerations

While demonstrating 100% plan validity and 80% grounding correctness on owner-run synthetic evaluations, the model exhibits a weak abstention rate (2/6 correct). This means it may incorrectly attempt to navigate when it should abstain. Therefore, it is not suitable for autonomous or high-stakes use without a robust external validating controller to prevent incorrect actions. The model's training data is proprietary and not publicly available for independent verification.

Finetunes
2 models