Xerv-AI/MAXWELL
Xerv-AI/MAXWELL is a 1.5 billion parameter Qwen2.5-Math-1.5B-Instruct variant developed by Xerv-AI, specialized for high-precision analytical reasoning, mathematical computation, and physics problem-solving. It utilizes a unique Systematic Temperature-Sweep Synthesis (STSS) inference framework to reduce hallucination rates by generating and aggregating responses across varying temperatures. The model supports a 4096-token context length via RoPE scaling and is deployed in a merged FP16 format after 4-bit quantization.
Loading preview...
MAXWELL: Specialized STEM Reasoning Model
MAXWELL, developed by Xerv-AI, is a fine-tuned variant of the Qwen2.5-Math-1.5B-Instruct architecture, specifically optimized for high-precision analytical reasoning, mathematical computation, and physics problem-solving. This 1.5 billion parameter model leverages a unique Systematic Temperature-Sweep Synthesis (STSS) inference framework to enhance accuracy and reduce hallucinations.
Key Capabilities & Features
- STSS Inference Framework: Replaces single-shot generation with a two-phase meta-reasoning protocol. It generates candidate responses across a spectrum of temperatures (0.1 to 0.9) and then re-prompts the model to aggregate these solutions at a low temperature (0.1) for logical cross-referencing and verification.
- Enhanced Analytical Reasoning: Demonstrates strong logical consistency in cognitive reflection tests, maintaining deterministic responses even at high temperatures.
- Optimized for STEM: Fine-tuned for mathematics and physics, with specific attention to empirical STEM constraints during aggregation.
- 4096-Token Context: Utilizes RoPE scaling to support a maximum context length of 4096 tokens, crucial for the multi-pass STSS process.
When to Use MAXWELL
MAXWELL is ideal for use cases requiring high-accuracy, verifiable solutions in STEM fields, particularly for complex mathematical problems, physics simulations, and analytical reasoning tasks where hallucination reduction is critical. Its STSS framework provides a robust method for ensuring logical consistency, making it suitable for applications where precision is paramount. However, users should note the 6x compute multiplier due to the multi-pass inference, which increases latency and cost compared to standard LLMs.