ZenithLLM/zen-ingress-1-3b-dpo-merged

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ZenithLLM's Zen Ingress 1-3B-DPO-Merged is a 3.1 billion parameter causal language model, fine-tuned from Qwen 2.5 3B. It is specifically engineered for rigorous step-by-step reasoning, clean algorithmic efficiency, and optimized for edge-device deployment. This model excels at providing direct, logical, and concise responses, suppressing conversational fluff and disclaimers. Its primary strength lies in mathematical deduction, clean coding style, and first-principles reasoning.

Loading preview...

Zen Ingress 1-3B-DPO-Merged Overview

ZenithLLM's Zen Ingress 1 is a 3.1 billion parameter causal language model, meticulously fine-tuned from Qwen 2.5 3B. It is designed with a "No-Compromise" strictness rule, focusing on rigorous step-by-step reasoning, clean algorithmic efficiency, and edge-device deployment.

Key Development Phases & Features

The model underwent a rigid 4-phase development cycle:

  • Data Generation & Curation: Involved 15,070 strictly filtered samples, ensuring high signal-to-noise ratio and direct logical paths.
  • Supervised Fine-Tuning (SFT): Trained for 3 epochs, with a checkpoint available at ZenithLLM/zen-ingress-1-3b-sft-merged.
  • Direct Preference Optimization (DPO) Alignment: Utilized a curated Multi-Domain Mixture of ~12,000 preference pairs, including data for arithmetic deduction, algorithmic efficiency, and suppression of conversational filler. This phase specifically targeted enhancing step-by-step reasoning, clean coding, and reducing hallucinations.
  • Extreme Edge Optimization: The model supports Imatrix Quantization for highly optimized GGUF versions, targeting mobile edge bandwidth with options like IQ4_NL for optimal balance and Q5_K_M for high fidelity.

Model Identity & Strengths

Zen Ingress 1-3B-DPO-Merged is characterized by its persona that heavily suppresses "As an AI..." disclaimers and introductory fluff. It prioritizes concise, strictly formatted, and algorithmically sound responses, making it ideal for applications requiring direct and logical outputs without verbosity.

Finetunes
1 models