Phonsiri/Qwen2.5-3B-SFT-opus-4.6-reasoning

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Feb 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Phonsiri/Qwen2.5-3B-SFT-opus-4.6-reasoning is a 3 billion parameter Supervised Fine-Tuned (SFT) model based on the Qwen2.5-3B architecture, developed by Phonsiri. It is specifically optimized for robust step-by-step reasoning, conversational intelligence, and structured formatting, trained on a dense, high-quality reasoning dataset. This model excels at logic puzzles and complex instruction-following, supporting both English and Thai languages with a 32768 token context length.

Loading preview...

Model Overview

Phonsiri/Qwen2.5-3B-SFT-opus-4.6-reasoning is a 3 billion parameter model, fine-tuned from the Qwen2.5-3B base, with a strong emphasis on developing advanced reasoning capabilities. This Supervised Fine-Tuned (SFT) model is designed to handle complex logical tasks and maintain conversational coherence, even on consumer-grade hardware, leveraging its compact size.

Key Capabilities

  • Enhanced Reasoning: Aggressively trained on a specialized dataset (nohurry/Opus-4.6-Reasoning-3000x-filtered) to imprint deep logic structures, enabling robust step-by-step reasoning.
  • Conversational Intelligence: Demonstrates strong conversational alignment and the ability to follow complex instructions.
  • Structured Output: Capable of generating structured responses, including the use of <think> tags for internal thought processes.
  • Multilingual Support: Supports both English and Thai languages.
  • Efficient Performance: Despite its compact 3 billion parameters, it performs effectively on logic puzzles and complex instruction-following.

Training Details

The model underwent full Supervised Fine-Tuning (SFT) using SFTTrainer with a context window of 8192 tokens during training to accommodate long reasoning traces. It utilizes bfloat16 precision and was trained for 3 epochs with an AdamW optimizer and a learning rate of 2.0e-5.

Good For

  • Applications requiring detailed, step-by-step logical reasoning.
  • Chatbots or assistants needing strong conversational intelligence and complex instruction adherence.
  • Scenarios where structured output, such as internal thought processes, is beneficial.
  • Deployment on consumer hardware due to its efficient 3 billion parameter size.

Finetunes
1 models