sriq-ai/Sriqwen-V1.3

SRIQwen-V1.3: Concise Reasoning from SRIQ

SRIQwen-V1.3 is a 27 billion parameter language model developed by SRIQ, built upon the Qwen/Qwen3.8-27B base model. Its primary differentiation lies in its supervised fine-tuning objective: to generate shorter reasoning traces in Simplified Chinese while preserving the integrity and quality of the final answer. This is achieved by training on a custom dataset where assistant targets pair compressed reasoning with original final answers.

Key Features and Training:

• Base Model: Qwen/Qwen3.8-27B.
• Fine-tuning Method: LoRA supervised fine-tuning, with the adapter merged directly into the released weights for simplified inference.
• Dataset: Trained on sriq-ai/sriq-sft-v1.3, which focuses on compressing Simplified Chinese reasoning steps.
• Context Length: While the base model supports 262,144 tokens, fine-tuning used sequences up to 131,072 tokens.
• Vision Capability: The model supports image inputs and can be served with its vision tower enabled, or disabled for language-only tasks.
• Reasoning Output: Reasoning steps are emitted within <think> tags in Simplified Chinese, followed by the final answer in the prompt's language.

Important Considerations:

• No Benchmarks: SRIQwen-V1.3 was released without official benchmarks. Users are advised to evaluate its performance on their specific workloads.
• Usage: The model can be deployed with vLLM, supporting vision inputs, or loaded via Hugging Face Transformers using AutoProcessor and AutoModelForImageTextToText.