S5-nano
Maritaca AI's 4B model for Brazilian Portuguese, built on Qwen3.5-4B, with Think and Instruct modes, 256K context and Apache 2.0.
In short. S5-nano is a small language model from the Brazilian company Maritaca AI that is tuned for Brazilian Portuguese and still handles tool calls and structured output. It can think before it answers or reply directly, and the 2.8 GB build runs on an ordinary computer. It is free under Apache 2.0, and the scores come from Maritaca.
Summary
S5-nano is the open-weight small model from Maritaca AI, a Brazilian company whose larger S5 is offered through its own chat and API. Maritaca built it on Qwen3.5-4B with continued pretraining followed by post-training, and the model card says it delivers substantial improvements on Brazilian benchmarks while keeping strong function calling, tool use and structured outputs. One checkpoint has two modes: Think, the default, reasons before answering, and Instruct answers directly, and you choose per request. The context window is up to 262,144 tokens, and the weights include a multi-token-prediction layer that vLLM can use for speculative decoding.
On Maritaca's own Observatório benchmarks, S5-nano in Think mode averages 75.25 against 68.68 for Qwen3.5-4B, 64.95 for Nemotron 3.5 Bolt E4B and 63.68 for Gemma 4 E4B, and in Instruct mode 66.51 against 61.84 for Qwen3.5-4B. The lead comes from tests written natively in Brazilian Portuguese, where it averages 83.74 against 72.49 for Qwen3.5-4B, with large gaps on OAB-Bench, Magis-Bench and the undergraduate part of MMLU-PT. In Think mode its translated-Portuguese and non-native averages are within a point of Qwen3.5-4B's and slightly lower, and it trails Qwen3.5-4B on IFBench, MATH-PT, LongBench v2 and Global-MMLU PT. These are the developer's own results, and we have not reproduced them.
The weights use Apache 2.0, the same license as the base model. The BF16 repository is 9.3 GB, the FP8 build is 5.8 GB and the GGUF builds run from 2.8 GB for Q4_K_M to 8.7 GB for BF16. The checkpoint still carries the base model's vision encoder, but Maritaca publishes text results only. The repositories first appeared on Hugging Face on 2 November 2026 and, in the 30 days to 11 November 2026, Hugging Face counted about 3,200 downloads for the GGUF repository, 1,200 for the BF16 weights and 300 for FP8.
Good for
- Assistants and tools that must write and understand Brazilian Portuguese on modest hardware, from the 2.8 GB Q4_K_M build to a single small GPU
- Professional and university-level questions in Brazilian Portuguese, where the card shows its largest gains over the base model on OAB-Bench, Magis-Bench and MMLU-PT undergraduate questions
- Function calling and structured JSON output in a 4B model, with Think or Instruct chosen for each request
Watch out for
- Every score is Maritaca's own, and in Think mode the model trails Qwen3.5-4B on IFBench, MATH-PT, LongBench v2 and Global-MMLU PT
- Outside native Brazilian Portuguese tests there is no gain, with 54.11 on the card's non-native average against 54.82 for Qwen3.5-4B in Think mode
- The larger S5 is offered through Maritaca's chat and API, and S5-nano is the model with open weights
- The checkpoint ships a vision encoder, but only text results are published, so treat image input as untested
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the S5-nano Q4_K_M (GGUF) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| S5-nano BF16 | 4.66B | 11 GB | 5.36 GB | 3.22 GB | Laptop or phone |
| S5-nano FP8 FP8 | 4.66B | 11 GB | 5.36 GB | 3.22 GB | Laptop or phone |
| S5-nano Q4_K_M (GGUF) INT4 | 4.66B | 11 GB | 5.36 GB | 3.08 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is S5-nano?
- S5-nano is a small language model from the Brazilian company Maritaca AI that is tuned for Brazilian Portuguese and still handles tool calls and structured output. It can think before it answers or reply directly, and the 2.8 GB build runs on an ordinary computer. It is free under Apache 2.0, and the scores come from Maritaca.
- Can I use S5-nano commercially?
- Yes. S5-nano is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is S5-nano open source?
- S5-nano is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run S5-nano on my own computer?
- Yes. The smallest variant, S5-nano Q4_K_M (GGUF), needs about 3.08 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was S5-nano released?
- S5-nano was released by Maritaca AI on November 2, 2026.
- How long a context window does S5-nano have?
- S5-nano supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- S5-nano model card (Hugging Face) huggingface.co
- Maritaca AI homepage maritaca.ai
- Observatório Maritaca, the benchmark results behind the model card maritaca-ai.github.io
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Nemotron 3.5 NVIDIA | 8B – 120B 12B active | 1M | OpenMDW-1.1 (Permissive) | 5.52 GB Laptop or phone | |
| LFM3 Liquid AI | 1.73B – 2.17B | 256K | LFM Open 1.0 (Conditional) | 1.19 GB Laptop or phone |