lightonai/Qwen3-8B-ES
The lightonai/Qwen3-8B-ES is an 8 billion parameter native reasoning model, fine-tuned from Qwen/Qwen3-8B-Base, specifically designed for Spanish language processing. It produces its entire chain-of-thought and final answer in Spanish, leveraging a 32,768 token context length. This model is optimized for complex reasoning tasks in Spanish, making it suitable for applications requiring detailed Spanish-language problem-solving.
Loading preview...
Overview
lightonai/Qwen3-8B-ES is an 8 billion parameter language model developed by LightOnAI, fine-tuned from Qwen/Qwen3-8B-Base. Its primary distinction is its native Spanish reasoning capability, generating both its chain-of-thought (CoT) and final answers entirely in Spanish. This model is part of a research effort to understand and address the multilingual reasoning gap, as detailed in the paper "Rethinking the Multilingual Reasoning Gap with Layer Swap".
Key Capabilities & Features
- Native Spanish Reasoning: Processes and generates reasoning traces and answers exclusively in Spanish.
- Extended Context Window: Supports a context length of 32,768 tokens, enabling processing of longer Spanish texts and complex problems.
- Specialized Training: Fine-tuned on approximately 10 billion tokens over 2 epochs using the Spanish split of the
lightonai/Dolci-Think-SFT-32B-Multilingualdataset.
Performance Highlights
Evaluated on Spanish versions of various benchmarks, Qwen3-8B-ES demonstrates strong performance in reasoning and language understanding:
- MGSM-Rev2: 93.20%
- Global-MMLU-Lite: 76.58%
- GPQA-Diamond: 55.15%
- AIME 24/25: 56.11%
- HumanEvalPlus: 81.00%
These scores indicate its proficiency across mathematical reasoning, general knowledge, and code generation tasks within the Spanish language context.
When to Use This Model
This model is ideal for applications requiring robust, native Spanish reasoning and problem-solving. It is particularly well-suited for tasks where a detailed, Spanish-language chain-of-thought is beneficial or required, such as educational tools, customer support in Spanish, or complex analytical tasks where the reasoning process needs to be transparent and in Spanish.