sglim/Qwen3-VL-8B-Thinking-SEA-Reasoning

VISIONPricing:Input $0.727 / Output $5.405Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 5, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

sglim/Qwen3-VL-8B-Thinking-SEA-Reasoning is an 8 billion parameter vision-language model developed by sglim, adapted from Qwen/Qwen3-VL-8B-Thinking. It is specifically fine-tuned for native multilingual Chain-of-Thought (CoT) reasoning in low-resource Southeast Asian (SEA) languages, utilizing the Onramp Sequence Cross-Distillation (OSCD) methodology. This model excels at providing reasoning traces in target SEA languages while delivering answers in English, making it suitable for applications requiring global interpretability of reasoning.

Loading preview...

Overview

sglim/Qwen3-VL-8B-Thinking-SEA-Reasoning is an 8 billion parameter vision-language model built upon the Qwen3-VL-8B-Thinking architecture. Its primary innovation lies in enabling native multilingual Chain-of-Thought (CoT) reasoning across various low-resource Southeast Asian (SEA) languages, including Filipino, Indonesian, Tamil, Thai, and Vietnamese, in addition to Chinese and English. This is achieved through a sequential post-training process using the Onramp Sequence Cross-Distillation (OSCD) methodology, which aligns reasoning traces to target SEA languages.

Key Capabilities

  • Multilingual CoT Reasoning: Supports deep, step-by-step reasoning in multiple SEA languages.
  • Vision-Language Integration: Inherits VL capabilities from the base Qwen3-VL model, allowing for image-based queries.
  • English Answer Blocks: This specific variant provides reasoning in SEA languages but outputs final answers in English, facilitating global interpretability.
  • Benchmark Performance: Demonstrates competitive performance on mathematical reasoning benchmarks like AIME25-SEA and HMMT25-SEA across supported languages.

Good For

  • Research in Multilingual LLMs: Particularly for studies focusing on low-resource languages and reasoning.
  • Applications Requiring Interpretable Reasoning: Ideal for scenarios where the reasoning process needs to be understood in a local SEA language, but the final output requires English for broader use.
  • Conversational AI Development: Can be integrated into systems needing native language interaction with English-based final responses.

Finetunes
2 models