Simonc-44/Cygnis-A3

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Cygnis-A3 V0.1 by Simonc-44 is a 7.6 billion parameter causal language model based on Qwen2.5-7B-Instruct, featuring a 32,768 token context length. This model is a linear weighted merge of Qwen2.5-Instruct, Qwen2.5-Coder, and Qwen2.5-Math, specifically optimized for code generation, mathematical problem-solving, and instruction following. It serves as a stable foundation, focusing on direct responses without explicit chain-of-thought reasoning, which is planned for future versions.

Loading preview...

Cygnis-A3 V0.1: A Merged Model for Code, Math, and Instructions

Cygnis-A3 V0.1 is the initial public release from Simonc-44, a 7.6 billion parameter language model built upon the Qwen2.5-7B-Instruct base. This version is constructed using a pure PyTorch linear weighted merge, combining the strengths of Qwen2.5-Instruct (70%), Qwen2.5-Coder (20%), and Qwen2.5-Math (10%). It features a substantial native context length of 32,768 tokens and operates in BFloat16 precision.

Key Capabilities

  • Code Generation: Enhanced capabilities for generating and understanding code, derived from the Qwen2.5-Coder component.
  • Mathematical Problem Solving: Improved performance on mathematical tasks, leveraging the Qwen2.5-Math integration.
  • Instruction Following: Robust ability to follow user instructions, inherited from the Qwen2.5-Instruct base.
  • Direct Responses: This V0.1 release provides direct answers without internal reasoning steps or Chain-of-Thought (CoT) capabilities, which are slated for integration in V0.2 with DeepSeek-R1.

Ideal Use Cases

  • Software Development: Generating Python functions, solving coding challenges, and assisting with programming tasks.
  • Technical Problem Solving: Addressing mathematical queries and problems requiring precise calculations.
  • General Instruction Following: Applications where clear, concise, and direct responses to instructions are preferred.

This model is designed as a stable and validated base, offering strong performance in its specialized domains before the introduction of more complex reasoning modules in subsequent iterations.

Finetunes
1 models