Cheva123/Jamjuri-EDGE-Preview-100
Cheva123/Jamjuri-EDGE-Preview-100 is a 4 billion parameter experimental model family based on Qwen3-4B-Instruct-2507, designed for local, offline, and tool-driven workloads. It excels in Thai and English, coding, calculation, instruction following, and IoT automation. This model uniquely offers multiple TIES-merged candidates and raw specialist adapters, allowing users to select or build models optimized for specific tasks like IoT, coding, or general instruction following.
Loading preview...
Jamjuri Edge: Compact 4B Models for Local AI
Jamjuri Edge is a family of experimental 4 billion parameter models developed by Chevalabs, built upon Qwen3-4B-Instruct-2507. It's specifically engineered for practical local AI applications, supporting both Thai and English. The core philosophy is to provide robust capabilities for scenarios where cloud models are impractical, such as offline systems, IoT devices, and private data environments.
Key Capabilities & Features
- Multilingual: Strong performance in both Thai and English.
- Specialized Tasks: Optimized for coding (Python), calculation, instruction following, and IoT/automation.
- Tool Calling: Designed for effective function calling and tool use.
- Modular Design: Released as a core model, raw specialist adapters (Control, Math, Code), and multiple TIES-merged candidates, enabling users to choose or customize for specific needs.
- Local Focus: Ideal for local, offline, and edge deployments.
Performance Highlights
The model family does not present a single "universal winner," but rather a set of candidates each excelling in different areas:
- TIES-452530: Leads in IoT (36.67) and HumanEval (84.76) benchmarks.
- TIES-402535: Achieves top scores in IFEval-TH (75.81) and GSM8K (93.48).
- TIES-352540: Strong in coding-heavy tasks, also scoring 84.76 on HumanEval.
Should I use this for my use case?
Jamjuri Edge is particularly suited for developers requiring compact, performant models for specific local or edge-based applications. If your project involves:
- IoT and Automation: Use TIES-452530 for local agents and tool-heavy workloads.
- Bilingual Instruction Following & General Use: TIES-402535 is balanced for Thai instruction and general technical tasks.
- Coding-Heavy Applications: TIES-352540 is optimized for programming experiments.
Its unique release strategy, offering both merged candidates and raw adapters, provides flexibility for researchers and developers to fine-tune or create custom merges for highly specialized requirements. The model is experimental, and users should be aware of potential trade-offs in performance across different capabilities due to its 4B parameter size.