valentin0901/Planium-4B
Planium-4B by valentin0901 is a 4.5 billion parameter fine-tune of Qwen3.5, specifically specialized in automated planning using PDDL (Planning Domain Definition Language). This model excels at translating plain-language requests into PDDL, explaining existing planning domains, and reasoning about states and goals for robotics and control systems. Its 4-bit quantized version is optimized to run efficiently on ordinary PCs or embedded boards like NVIDIA Jetson, enabling on-device operation without cloud APIs. It also features tool-calling capabilities for integration into agentic loops, driving parsers, planners, or plan validators.
Loading preview...
Overview
Planium-4B is a 4.5 billion parameter model developed by valentin0901, fine-tuned from Qwen3.5. Its core specialization lies in automated planning using the PDDL (Planning Domain Definition Language), making it particularly suitable for applications in robotics and other control systems. A key design goal was to enable on-device execution; its 4-bit quantized build is small enough to run on standard PCs or embedded systems like NVIDIA Jetson, eliminating the need for cloud APIs.
Key Capabilities
- PDDL Generation: Converts plain-language requests into PDDL, typically for defining goals or problem files within a given domain.
- PDDL Interpretation: Reads and explains existing PDDL domains, problems, or plans.
- Reasoning: Performs reasoning about states, preconditions, effects, and goals within planning contexts.
- Error Detection: Identifies issues and inconsistencies in PDDL code.
- Tool Calling: Trained to operate within agentic loops, capable of emitting tool calls to drive external parsers, planners, or plan validators.
- Efficiency: Mitigates a common failure mode of its base model (Qwen3.5) by keeping reasoning concise and committing to answers, avoiding prolonged, unproductive computations.
Limitations
While strong in classical planning (especially :strips and :strips :typing fragments), Planium-4B offers much less support for temporal planning, numeric fluents, and richer ADL constructs.
Good For
- Developers building neuro-symbolic planning systems that bridge LLMs with symbolic planners.
- Applications requiring on-device PDDL generation and interpretation for robotics or control systems.
- Scenarios where efficient, concise planning reasoning is critical.
Planium-4B is the underlying model for NeuroPlan, a system designed to translate requests into PDDL, execute plans, and manage the planning process.