wdrones/nemo-qwen3_5_4b_base_jellyfish
The wdrones/nemo-qwen3_5_4b_base_jellyfish model is an experimental 4.5 billion parameter vision-language model developed by NVIDIA, based on the Qwen3.5 architecture. It is fine-tuned for tool-calling and function-calling, accepting interleaved text and images as input to generate text outputs, including structured tool calls. This intermediate checkpoint (step 20500) is designed for research and experimentation in multimodal tool-use scenarios.
Loading preview...
Model Overview
This is an experimental, intermediate checkpoint (step 20500) from NVIDIA's Nemotron "edge" experiments, codenamed steep_jellyfish. It is a vision-language, tool-calling model built upon the Qwen3.5 architecture, specifically fine-tuned for tool-calling/function-calling (SFT).
Key Capabilities
- Multimodal Input: Accepts interleaved text and images (and video frames) as input.
- Tool Calling: Generates structured
<tool_call>blocks in its text output when tools are supplied, following a specific XML-like format. - Qwen3.5 Architecture: Based on the
Qwen3_5ForConditionalGenerationmodel, featuring a hybrid attention pattern and Grouped-Query Attention (GQA). - Large Context Window: Supports a maximum context length of up to 262,144 tokens.
- Experimental Nature: This is an early checkpoint intended for research and experimentation; behavior and formats may change.
Good For
- Research and Development: Ideal for exploring multimodal tool-calling and function-calling capabilities.
- Vision-Language Tasks: Handling inputs that combine both visual and textual information.
- Prototyping Tool-Integrated LLM Applications: Experimenting with how LLMs can interact with external functions or APIs.
Limitations
As an early, intermediate checkpoint, this model has not undergone full safety or capability evaluation. Its outputs may be inconsistent, and no benchmark results are published. It is not recommended for production use without further evaluation and refinement.