ZENLLC/HUX-1
ZENLLC/HUX-1 is a 4 billion parameter causal language model developed by ZEN AI Co., built upon Qwen3-4B-Instruct-2507. It features a native context length of 262,144 tokens and is specifically optimized for AI agent generation, prompt enhancement, structured system outputs, and workflow automation logic. This model excels in instruction following, logical reasoning, and tool usage, demonstrating significant improvements across general capabilities including mathematics, science, and coding.
Loading preview...
ZENLLC/HUX-1: An Optimized Qwen3-4B Variant
HUX-1 is a 4 billion parameter causal language model developed by ZEN AI Co., serving as the internal intelligence layer for their Arsenal platform. It is built on the enhanced Qwen3-4B-Instruct-2507 base model, which features a native context length of 262,144 tokens.
Key Capabilities & Enhancements
This model demonstrates significant improvements in general capabilities, including:
- Instruction Following & Logical Reasoning: Enhanced ability to understand and execute complex instructions.
- Text Comprehension & Mathematics: Stronger performance in understanding diverse texts and solving mathematical problems.
- Coding & Tool Usage: Improved capabilities in code generation and effective utilization of external tools, making it suitable for agentic workflows.
- Long-tail Knowledge & Multilingualism: Substantial gains in covering less common knowledge across multiple languages.
- User Alignment: Markedly better alignment with user preferences for subjective and open-ended tasks, leading to more helpful and higher-quality text generation.
Optimized Use Cases
HUX-1 is specifically optimized for:
- AI Agent Generation: Creating and managing AI agents.
- Prompt Enhancement: Refining and improving prompts for better model outputs.
- Structured System Outputs: Generating outputs in predefined formats.
- Workflow Automation Logic: Developing logic for automated processes.
Notably, this model operates in a "non-thinking mode," meaning it does not generate <think></think> blocks, simplifying its output structure. Benchmarks show HUX-1 (as Qwen3-4B-Instruct-2507) outperforming previous Qwen3-4B versions and even larger models like Qwen3-30B-A3B Non-Thinking in several categories, including MMLU-Pro, GPQA, AIME25, HMMT25, ZebraLogic, LiveBench, Creative Writing, and various agentic tasks (BFCL-v3, TAU1-Retail, TAU1-Airline, TAU2-Retail, TAU2-Airline).