S5-nano: A small Portuguese-language model by Maritaca in Brazil

A 4B model developed by Maritaca AI in Brazil through continued pre-training and post-training on Qwen3.5-4B, primarily targeting Brazilian Portuguese while retaining function calling, tool use, and structured output capabilities. It offers two modes: thinking (default) and direct answer, with a context window of 262,144 tokens. The weights include MTP layers for speculative decoding. Officially, the model achieves an average score of 75.25 in thinking mode, compared to 68.68 for Qwen3.5-4B, with improvements mainly driven by local Brazilian tests; tests translated into Portuguese and non-Portuguese tests are similar to or slightly lower than the base model. Three versions are provided: BF16 (9.3 GB), FP8 (5.8 GB), and GGUF (starting from Q4_K_M 2.8 GB).

Original title: maritaca-ai/S5-nano

A 4B model developed by Maritaca AI in Brazil through continued pre-training and post-training on Qwen3.5-4B, primarily targeting Brazilian Portuguese while retaining function calling, tool use, and structured output capabilities. It offers two modes: thinking (default) and direct answer, with a context window of 262,144 tokens. The weights include MTP layers for speculative decoding. Officially, the model achieves an average score of 75.25 in thinking mode, compared to 68.68 for Qwen3.5-4B, with improvements mainly driven by local Brazilian tests; tests translated into Portuguese and non-Portuguese tests are similar to or slightly lower than the base model. Three versions are provided: BF16 (9.3 GB), FP8 (5.8 GB), and GGUF (starting from Q4_K_M 2.8 GB).

Why it matters A sample of a small model localized for non-English languages: significant improvements in local Brazilian tests, retained tool-calling capabilities, and the GGUF version runs on laptops

Upvotes 71 · Likes 72 · Downloads 1099

Reasons

  • Official evaluations are categorized into local Brazilian, Portuguese-translated, and non-Portuguese tests, making it clear where the improvements lie
  • Apache-2.0 license, with ready-to-use commands for vLLM, SGLang, and llama.cpp

Usage

  • vllm serve maritaca-ai/S5-nano --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder; switch modes using enable_thinking in chat_template_kwargs
  • Local: llama-server -hf maritaca-ai/S5-nano-GGUF:Q4_K_M --jinja -c 32768

Risks

  • Primarily targeted at Portuguese, with no mention of Chinese capabilities in the model card
  • Evaluations are self-reported by the official team
  • This entry has not been tested; refer to the model card for performance details

Open

Hugging Face Trending

Explore the latest AI Radar updates