RandomFrontlines/Orenis-3B-Light-Max

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Orenis-3B-Light-Max is a 3.1 billion parameter causal language model developed by OrenCraft Labs, fine-tuned on Qwen2.5-3B-Instruct. It features 100% uncompressed FP16 precision and excels at strict instruction following, anti-sycophancy, and structured reasoning. This model is specifically designed for use cases requiring precise adherence to rules, such as JSON schema generation, and includes an autonomous web search grounding protocol.

Loading preview...

Orenis-3B-Light-Max: Precision and Grounded Reasoning

Orenis-3B-Light-Max, developed by OrenCraft Labs (founded by Lee Zinu), is a 3.1 billion parameter model built upon Qwen2.5-3B-Instruct. This flagship model of the Light tier emphasizes uncompressed FP16 precision, ensuring full mathematical fidelity and robust code generation logic. It is engineered for stringent instruction adherence, anti-sycophancy, and structured reasoning, making it distinct in its ability to follow complex rules and avoid validating false premises.

Key Capabilities

  • 100% Uncompressed Precision (FP16): Maintains full mathematical fidelity and code generation capabilities.
  • IFEval Constraint Adherence: Strictly follows negative rules, exact sentence limits, and JSON schemas, achieving 60.19% on Instruction-Level Rule Adherence and 48.43% on Strict Multi-Constraint Following in IFEval benchmarks.
  • Anti-Sycophancy: Designed to challenge erroneous math, false premises, and common myths, promoting truthful and objective responses.
  • Autonomous Tool Protocol: Natively emits <search>query</search> tags when real-time information is required, enabling integration with live web search tools like DDGS for grounded responses.

Performance Highlights

Benchmarked in FP16 with a ChatML template, Orenis-3B-Light-Max achieved 64.22% on GSM8K (5-shot) for grade school math reasoning. Its strong performance in IFEval highlights its capability in instruction-level rule adherence and strict multi-constraint following.

When to Use This Model

This model is ideal for applications demanding high precision, strict adherence to complex instructions, and factual accuracy. Its autonomous search capability makes it suitable for tasks requiring up-to-date information, while its anti-sycophancy features ensure reliable and objective outputs. Developers can leverage its ability to generate structured outputs like JSON and follow specific formatting rules, making it a strong candidate for automated workflows and precise data generation.

Finetunes
1 models