hungrynovalabs/nova-pup-4b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 30, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Nova Pup 4B by Hungry Nova Labs LLC is a 4.5 billion parameter language model based on the InternScience Agents-A1-4B (Qwen3.5 hybrid) architecture. It is specifically fine-tuned as a Linux systems specialist and multi-agent problem solver, demonstrating significant improvement on Linux diagnostics exams. This model is optimized for agent swarms and provides step-by-step reasoning with a distinct "toon pup" persona.

Loading preview...

Nova Pup 4B: A Linux Systems Specialist

Nova Pup 4B, developed by Matthew Salinas Hernandez of Hungry Nova Labs LLC, is a 4.5 billion parameter model built upon the InternScience Agents-A1-4B (Qwen3.5 hybrid linear-attention) base. It is uniquely designed as a Linux systems specialist and multi-agent problem solver, trained end-to-end on a single RTX 5090.

Key Capabilities & Differentiators

  • Linux Expertise: Achieves a 75% score on a sealed closed-book Linux diagnostics exam, a 3x improvement over its base model (25%).
  • Specialization without Forgetting: Maintains strong general reasoning with MMLU at 73.7% (vs 74.0% for base), indicating effective specialization without catastrophic forgetting.
  • Multi-Agent Design: Built for agent swarms, allowing an 8-agent pack to run efficiently on a single RTX 5090 (15.5 GB VRAM, 1,539 tok/s aggregate via Ollama).
  • Rigorous Reasoning: Provides playful, rigorous "toon pup" persona with step-by-step reasoning and machine-checkable final answers.
  • Training Methodology: Underwent two phases of training: continued pretraining on a curated Linux corpus and Solver SFT on a programmatically generated puzzle curriculum including logic-grid deduction and bash-pipeline reconstruction.

Use Cases & Considerations

Nova Pup 4B is ideal for local Linux study, troubleshooting assistance, explaining concepts, and applying Linux knowledge to new problems. While it excels in its specialized domain, users should note a regression in mathematical (GSM8K dropped 6.7 points) and TruthfulQA (dipped 3.7 points) performance compared to its base model. It is designed to be small, fast, and replicable in swarms, rather than out-reasoning frontier models in general tasks. The model has an 8k context in its published configuration.

Finetunes
3 models