abeja/ABEJA-Qwen3-14B-Agentic-256k-v0.1
ABEJA-Qwen3-14B-Agentic-256k-v0.1 is a 14 billion parameter language model developed by ABEJA, based on Alibaba's Qwen3-14B. It features an extended context length of 256k tokens and is specifically fine-tuned to enhance agentic capabilities such as planning and tool use. This model is primarily designed for agent-based applications, focusing on iterative thought and tool interaction rather than broad general-purpose use.
Loading preview...
ABEJA-Qwen3-14B-Agentic-256k-v0.1: Agentic LLM with 256k Context
ABEJA-Qwen3-14B-Agentic-256k-v0.1 is a specialized large language model developed by ABEJA, building upon Alibaba's Qwen3-14B. This 14 billion parameter model has undergone additional training to significantly improve its performance in agentic scenarios.
Key Capabilities & Features
- Extended Context Length: Supports an impressive 256k token context window, enabling the processing of extensive information for complex tasks.
- Enhanced Agentic Abilities: Specifically fine-tuned to excel in planning and tool use, facilitating robust agent-based applications.
- Iterative Thought and Tool Interaction: Designed to support continuous loops of reasoning and external tool integration, crucial for autonomous agents.
Primary Use Case
This model is primarily intended for agent utilization, where its long context and agentic capabilities can be leveraged for sophisticated task execution. It is not optimized for broad general-purpose applications but rather for specific agent-driven workflows.
Recommended Usage
Similar to its base model, Qwen3-14B, ABEJA recommends specific generation parameters for optimal performance:
Temperature=0.6TopP=0.95TopK=20MinP=0
It is advised to avoid greedy decoding to prevent performance degradation and repetitive outputs. Further details on the development process, including reinforcement learning techniques, are available in ABEJA's technical blogs.