agentlans/Qwen3.5-2B-Instruct-SingleTurn

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

agentlans/Qwen3.5-2B-Instruct-SingleTurn is a 2.3 billion parameter language model fine-tuned from Qwen/Qwen3.5-2B, specifically optimized for single-turn English instruction-following tasks. It suppresses prolonged Chain-of-Thought reasoning and reduces Chinese text generation, making it suitable for lightweight deployment in English-only environments. This model excels in direct instruction, programming, and mathematics, but is not designed for multi-turn conversations or creative writing.

Loading preview...

Model Overview

agentlans/Qwen3.5-2B-Instruct-SingleTurn is a specialized 2.3 billion parameter model derived from the Qwen/Qwen3.5-2B base. Its primary distinction lies in its optimization for single-turn English instruction-following. This fine-tuning process specifically addresses and mitigates the base model's tendency to generate extensive Chain-of-Thought (CoT) reasoning and unwanted Chinese text, making it more direct and efficient for specific use cases.

Key Capabilities & Features

  • Single-Turn Focus: Designed exclusively for processing and responding to individual instructions without conversational context.
  • English-Centric: Optimized to primarily generate English responses, reducing multilingual interference.
  • Compact & Efficient: At 2.3 billion parameters, it's suitable for lightweight deployment and downstream fine-tuning.
  • Targeted Domains: Best suited for tasks in English instruction, programming, and mathematics.

Training Insights

The model was fine-tuned using the agentlans/sft-data dataset, employing packed sequences with a 2048 token cutoff. It leveraged FlashAttention-2 for efficiency, NEFTune regularization, and rsLoRA for PEFT, with a rank of 16 and alpha of 32.

Intended Use Cases

  • Direct Instruction Following: Ideal for applications requiring concise, single-response answers to prompts.
  • Lightweight Deployment: Its smaller size makes it suitable for environments with limited computational resources.
  • Specialized Tasks: Effective for programming assistance and mathematical problem-solving where direct answers are preferred.

Limitations

Due to its compact size, the model may exhibit occasional factual inaccuracies or basic reasoning errors. It is not recommended for multi-turn conversations, roleplay, or creative writing tasks, and extended responses might become repetitive.