S5-nano

S5-nano

Maritaca AI's small Portuguese and English model built on Qwen3.5-4B, with Think and Instruct modes and strong tool use, in BF16 for vLLM and SGLang. Suits developers building Brazilian Portuguese assistants.

Made by maritaca-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: S5-nano is a small model from Maritaca AI, built on Qwen3.5-4B through continued pretraining and then post-training. It reads images and text, writes text, and works in Portuguese and English. The card says it improves on Brazilian benchmarks while staying good at function calling, tool use and structured outputs. This repository holds the original BF16 weights, 9.3 GB.

Who it suits: Developers who build assistants for Brazilian Portuguese and teams that need a small model able to call tools and return structured output.

Industry | Who uses it | How they use it | Example
Software | Developers in Brazil | Build chat and agent features in Portuguese | Answer customer questions in Portuguese
Operations | Automation teams | Have the model call functions and tools | Return structured output for another system
Education | Students | Study the Think and Instruct modes of a small model | Compare both modes on one question
Research | Language researchers | Test a model on Brazilian benchmarks | Run it against their own test set

Skip it if you need a huge model, because this one has about 4B parameters.

Where to use it:

• vLLM and SGLang, for this BF16 version
• The FP8 version, on GPUs with native FP8 support such as Ada, Hopper and Blackwell
• llama.cpp and compatible apps, through the GGUF versions

Use case | Who | What they get back
Portuguese chat | Assistant builders | Replies written in Brazilian Portuguese
Tool calling | Developers | Function calls and structured output
Direct answers | Apps that need short replies | Instruct mode answers without a reasoning step

How to set it up:

• Python and the Hugging Face tools
• A serving engine such as vLLM or SGLang for this version
• About 9.3 GB of disk for the BF16 weights

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Serve them with vLLM or SGLang. Pick the FP8 or GGUF version if your hardware suits it better.

pip install -U huggingface_hub
hf download maritaca-ai/S5-nano

How to use it: Choose a mode. Think is the default and reasons before it answers. Instruct answers straight away. The card text here shows no serving command.

Tool | What it does
Think mode | Reasons before answering, and is the default
Instruct mode | Answers directly with no reasoning step
MTP layer | Supports speculative decoding
hf download | Copies the repository files to your computer

Example: Ask it, in Portuguese, to explain a short paragraph in simple words. In Think mode it reasons first and then answers. In Instruct mode it goes straight to the answer.

For newcomers and experts: If you are new to this, use Instruct mode first for quick answers, then try Think on a harder question and compare.

If you know the field, the FP8 copy stores linear layer weights in E4M3, with a scale for each output channel. It also quantizes activations for every token at run time. BF16 precision is kept for the MTP layer, the vision encoder, the embeddings and a few small projections.

You send text or an image
Think mode reasons first
Or Instruct mode answers directly
The model writes the reply

Title: Weight size by version
BF16: 9.3 GB
FP8: 5.8 GB
GGUF BF16: 8.7 GB
GGUF Q8_0: 4.6 GB
GGUF Q4_K_M: 2.8 GB

Good to know:

• The context window reaches 262,144 tokens.
• The card says results are published on Maritaca's own results site, and its text here lists no scores.
• Hugging Face counts 4.7B parameters, while the card calls it a 4B model.
• License: apache-2.0.
• Cost to run: none named, only your own hardware.
• Data it reaches: the text and images you give it. The source names no account or outside service.
• Not for: the source names none.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.7B
• Weights: about 9.3 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,357
• Likes: 77
• License: apache-2.0
• Languages: pt, en
• Quantized versions: 9
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, conversational

Links

• Model: https://huggingface.co/maritaca-ai/S5-nano
• Files: https://huggingface.co/maritaca-ai/S5-nano/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download maritaca-ai/S5-nano

Then follow the usage in its model card. Nothing in this kit runs by itself.