togethercomputer/HybridAgent-14B-Agentic
HybridAgent-14B-Agentic by togethercomputer is a 14 billion parameter LLM agent built on Qwen3-14B-Base, designed for agentic reasoning tasks. It features a unique HybridLink that enables dynamic navigation between latent reasoning and text-based interaction within a single generation process, allowing it to "think" in a continuous latent space. This model excels at complex problem-solving by integrating a Python interpreter and web search tool, making its latent thoughts interpretable as pseudo-text.
Loading preview...
Overview of HybridAgent-14B-Agentic
HybridAgent-14B-Agentic, developed by togethercomputer, is a 14 billion parameter language model agent specifically engineered for agentic reasoning tasks. It is built upon the Qwen3-14B-Base architecture and incorporates a novel "HybridLink" mechanism. This allows the agent to dynamically switch between continuous latent space reasoning and text-based interaction during a single generation process.
Key Capabilities
- Hybrid Reasoning: Integrates latent (thought-based) and text-based (action-based) reasoning within a unified process.
- Interpretable Latent Mode: Latent thoughts are directly interpretable, presented as pseudo-text, offering transparency into the agent's decision-making.
- Tool Integration: Can call external tools like a Python interpreter and a web search tool across multiple turns, leveraging a shared working memory.
- Dynamic Navigation: Efficiently navigates between internal thought processes and external interactions to solve complex problems.
Good For
- Developing advanced AI agents requiring complex, multi-step reasoning.
- Applications needing transparent and interpretable agent behavior.
- Tasks that benefit from dynamic tool use, such as code execution and information retrieval.
- Research into hybrid reasoning architectures for large language models.
This model utilizes HybridSFT followed by HybridRL for its training, enhancing its performance in agentic scenarios. The architecture includes standard Transformers files for the backbone, along with specific HybridLink weights (adapter.pt) for its unique functionality. For full hybrid generation capabilities, the dedicated HybridAgent code is required.