maoge888888/HaS_Text_0209_0.6B

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 26, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

The HaS Text Model (maoge888888/HaS_Text_0209_0.6B) is a 0.6 billion parameter, full-precision (FP16) privacy model designed for on-device entity recognition, anonymization, and restoration. It provides a complete agentic pipeline for handling sensitive data locally, supporting 8 languages natively including Chinese, English, and Japanese. This model excels at maintaining data usability post-anonymization through structured semantic tags and coreference resolution, making it ideal for privacy-preserving data processing without data leaving the device.

Loading preview...

HaS Text Model: On-Device Privacy Pipeline

The HaS (Hide and Seek) Text Model is a 0.6 billion parameter, full-precision (FP16) model developed by maoge888888, specifically engineered for on-device privacy. Unlike traditional anonymization methods that rely on simple pattern matching, HaS offers a comprehensive agentic privacy pipeline that ensures data never leaves the device, supporting local inference without network requirements. It natively handles 8 languages, including Chinese, English, Portuguese, French, Spanish, German, Korean, and Japanese.

Key Capabilities

  • Advanced Anonymization: Utilizes 3-Level Semantic Tags (e.g., <Amount[1].ContractAmount.NumberSymbol>) instead of generic redactions, preserving data usability for downstream LLMs.
  • Coreference Resolution: Automatically unifies different forms of the same entity (e.g., "CloudGenius Inc.", "CloudGenius") under a single ID, ensuring logical coherence across text.
  • Multi-turn Consistency: Maintains consistent entity IDs across multiple turns or long documents through historical mapping dictionaries.
  • Reversible Restoration: Allows anonymized text to be processed by cloud LLMs, with the ability to restore original values using the 'Seek' capability.
  • Open-set Entity Types: Trained on approximately 70,000 entity types, enabling users to define custom entity types without predefined limitations.
  • On-Device Performance: Achieves high inference speeds on Apple M4 Pro, with prefill rates up to 4,904 tok/s (llama.cpp) and decode rates up to 156 tok/s (mlx_lm) for the FP16 version.

Good for

  • Privacy-Preserving Data Sharing: Automatically anonymize files, emails, or code before sharing, with the option to restore original content.
  • Secure Cloud LLM Interaction: Anonymize sensitive information before sending to cloud-based LLMs and restore responses for privacy.
  • Building Privacy Knowledge Bases: Ingest documents into knowledge bases in an anonymized form, restoring content only when needed.
  • Log Redaction: Batch-anonymize operational logs to protect sensitive information before analysis or sharing with support teams.
  • AI Agent Memory Privacy: Store long-term agent memories in an anonymized format, restoring them on demand for secure recall.