Not tested by us yet. S5-nano ships as downloadable weights and no inference provider serves it, so we could not call it through our harness. Everything below comes from the model card on Hugging Face. We'll update this post when a hosted endpoint appears and we can run it against Qwen3.5-4B on our own Portuguese prompts.
S5-nano is a roughly 4B-parameter open-weight vision-language model from Maritaca AI, a Brazilian lab, released under Apache-2.0 with a 262,144-token context window. On Maritaca's own 23-benchmark suite it scores 75.25 in Think mode against 68.68 for Qwen3.5-4B. Almost all of that lead comes from Brazil-specific tasks. On translated Portuguese and non-native tests it is level with or slightly behind Qwen3.5-4B.
| S5-nano | |
|---|---|
| Maker | Maritaca AI (Brazil) |
| Size | About 4.2B text parameters per the card's table (Safetensors metadata shows 5B) |
| Architecture tag | qwen3_5 |
| License | Apache-2.0 |
| Context window | 262,144 tokens |
| Modalities | Image and text in, text out |
| Languages | Portuguese and English |
| Modes | Think (default) and Instruct, switched per request |
| Extras | Multi-token prediction layer for speculative decoding |
| Hosted API | None found; no Inference Provider lists it |
What you get
The card ships three builds: BF16 at 8.7 GB for vLLM and SGLang, FP8 at 5.7 GB for GPUs with native FP8 (Ada, Hopper, Blackwell), and GGUF files for llama.cpp from 2.8 GB (Q4_K_M) to 8.7 GB, plus a 0.7 GB vision projector. A single server handles both modes: Think reasons first, Instruct answers directly with enable_thinking set to false.
The card gives different sampling settings for each mode. Think uses temperature 1.0, top-p 0.95 and presence penalty 1.5. Instruct uses temperature 0.7 and top-p 0.8. It also warns that the shipped generation_config.json lacks the presence penalty, so you must send it with every request. That is the sort of detail that quietly ruins a first evaluation if you miss it.
The card says tool calling is not supported in its evaluations, even though the vLLM command it gives enables a tool-call parser. Treat agent use as unproven.
The vendor-reported numbers
All scores below are from the model card and come from Maritaca's own evaluation suite (Observatório Maritaca). Each benchmark's primary metric is weighted equally, and all five comparison models were run on 23 of 23 tests.
| Group (Think mode) | S5-nano | Qwen3.5-4B |
|---|---|---|
| Native Brazilian tasks | 83.74 | 72.49 |
| Translated Portuguese | 86.94 | 87.61 |
| Non-native | 54.11 | 54.82 |
| Total | 75.25 | 68.68 |
Other Think-mode results on the card, S5-nano against Qwen3.5-4B:
- OAB-Bench (Brazilian bar exam, scored out of 10): 82.00 vs 45.88
- Prosa-PT (rubrics passed): 75.07 vs 56.89
- BrowseComp-Plus: 37.00 vs 27.00
- MATH-PT: 87.77 vs 91.07
- IFBench (strict instruction following): 42.67 vs 49.67
- LongBench v2: 49.70 vs 54.08
- HarmBench (safety, normalized): 91.25 vs 96.56
In Instruct mode the total is 66.51 against 61.84 for Qwen3.5-4B. The card also lists Nemotron 3.5 Bolt E4B (64.95 in Think mode), Gemma 4 E4B (63.68) and LFM2.5-24B-A2B (67.06) as comparisons, so S5-nano leads all of them on this suite.
Read the table with care. Maritaca built the benchmark suite, and the "native Brazilian" group is the one where it wins by 11 points. That group is also the model's reason to exist, so the gap is believable, but it is a claim about Portuguese for Brazil, not about small models in general. S5-nano loses to Qwen3.5-4B on instruction following, math, long context and harmful-prompt refusal, and the card says as much. If you need strict output formats or safety behavior, test those before switching. Evaluations also gave Think mode the remaining context for output with no cap, which makes Think scores expensive in tokens.
How it compares on paper
You can't rent S5-nano, so the nearest hosted comparison is a bigger Qwen. On the OpenRouter model list, Qwen3.5-9B costs $0.17 per million input tokens and $0.25 per million output tokens with a 262,144-token window. Qwen3.5-4B itself isn't listed there. S5-nano's draw is different: at 2.8 GB quantized, it fits on a laptop or a modest GPU, so the cost is hardware you already own and a serving setup you maintain. For what that involves, our guide to open-source LLMs you can run on-prem covers the trade-offs, and Mirage is a useful contrast: a vision model that is far smaller still.
The serving commands on the card are plain. vLLM takes vllm serve maritaca-ai/S5-nano --max-model-len 262144 --reasoning-parser qwen3, llama.cpp takes llama-server -hf maritaca-ai/S5-nano-GGUF:Q4_K_M --jinja, and Docker Model Runner takes docker model run hf.co/maritaca-ai/S5-nano. The card lists no hardware requirement. It only suggests --max-model-len 32768 for GPUs with less memory.
Who should care
- Teams building Portuguese products (legal, education, customer support, health) that want an on-device or self-hosted model and can afford to check the Brazilian-task claims on their own documents.
- Anyone who needs a small multimodal model under Apache-2.0 with a long context window and no usage restrictions.
- Everyone else can pass. For English-first work, general instruction following or tool use, the card's own numbers show Qwen3.5-4B equal or better.
The release date is not stated on the card, and Hugging Face showed 263 downloads in the last month when we read it, so the community has barely touched it. Independent confirmation of the Maritaca numbers does not exist yet.
What we'll test when we can
Once a hosted endpoint exists, we'll run S5-nano against Qwen3.5-4B on 10 to 15 Portuguese prompts (contract summaries, exam-style questions, customer emails) plus an instruction-following set, with judge answer order swapped, and report latency and cost alongside the results.