---
type: "post"
id: "maritaca-ai-open-weights-s5-nano-4b-qwen3-919e38"
slug: "maritaca-ai-open-weights-s5-nano-4b-qwen3-919e38"
title: "Maritaca AI open-weights S5-nano 4B (Qwen3.5-4B based) for Portuguese"
description: "Maritaca AI has released S5-nano, an open-weight model of about 4 billion parameters with separate reasoning and direct-answer modes, a context window of up to 262,144 tokens, and builds for several local-serving stacks."
retrieval_nugget: "Maritaca AI has released S5-nano, an open-weight model of about 4 billion parameters with separate reasoning and direct-answer modes, a context window of up to 262,144 tokens, and builds for several local-serving stacks. Its published Portuguese evaluations make it a candidate for teams testing self-hosted language workloads, though the results are Maritaca's own and need validation on real applications."
published_at: "2026-11-06"
updated_at: "2026-11-06"
record_date: "2026-11-06"
date_kind: "published_at"
topics: ["open-models","local-models","models"]
entities: ["Maritaca AI","S5-nano","Qwen3.5-4B","Hugging Face","vLLM","SGLang","llama.cpp"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Maritaca AI released S5-nano, a 4B Portuguese-focused open-weight model with Think and Instruct modes, a 262,144-token context window and downloadable BF16, FP8 and GGUF variants."
prior_refs: []
source_url: "https://huggingface.co/maritaca-ai/S5-nano"
source_urls: ["https://huggingface.co/maritaca-ai/S5-nano"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:maritaca-ai-open-weights-s5-nano-4b-qwen3-919e38"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/maritaca-ai-open-weights-s5-nano-4b-qwen3-919e38/","markdown":"https://newruntime.com/posts/maritaca-ai-open-weights-s5-nano-4b-qwen3-919e38.md","json":"https://newruntime.com/posts/maritaca-ai-open-weights-s5-nano-4b-qwen3-919e38.json"}
markdown_url: "https://newruntime.com/posts/maritaca-ai-open-weights-s5-nano-4b-qwen3-919e38.md"
---

# Maritaca AI open-weights S5-nano 4B (Qwen3.5-4B based) for Portuguese

## Retrieval answer

Maritaca AI has released S5-nano, an open-weight model of about 4 billion parameters with separate reasoning and direct-answer modes, a context window of up to 262,144 tokens, and builds for several local-serving stacks. Its published Portuguese evaluations make it a candidate for teams testing self-hosted language workloads, though the results are Maritaca's own and need validation on real applications.

## Summary
Maritaca AI has released S5-nano, an open-weight model of about 4 billion parameters with separate reasoning and direct-answer modes, a context window of up to 262,144 tokens, and builds for several local-serving stacks. Its published Portuguese evaluations make it a candidate for teams testing self-hosted language workloads, though the results are Maritaca's own and need validation on real applications.

## What happened
Maritaca AI's Hugging Face model card describes S5-nano as a 4B-parameter model with two modes. Think is the default and reasons before answering; Instruct answers directly without a reasoning step. The model supports up to 262,144 context tokens, and its weights include a multi-token prediction layer intended for speculative decoding.

The card lists a BF16 version at 8.7 GB for vLLM and SGLang, an FP8 version at 5.7 GB for those same serving frameworks, and a GGUF repository with files from 2.8 GB to 8.7 GB for llama.cpp. The GGUF options are BF16, Q8_0 and Q4_K_M; the repository also provides a 0.7 GB vision projector. Maritaca recommends FP8 for GPUs with native FP8 support, naming Ada, Hopper and Blackwell. The company also documents server examples, including vLLM's OpenAI-compatible endpoint, and a setting that enables speculative decoding with three speculative tokens. No hosted inference provider is listed as deploying the model on the card.

The model card publishes results from Maritaca's Observatório Maritaca for the original BF16 checkpoint on text tasks. Across 23 evaluations, with each primary metric weighted equally, S5-nano's average is 75.25 in Think and 66.51 in Instruct. The card lists Qwen3.5-4B at 68.68 and 61.84 in the corresponding modes. For native Brazilian Portuguese, the stated averages are 83.74 and 74.65 for S5-nano, versus 72.49 and 66.22 for Qwen3.5-4B. These are not wins on every task: for example, the card reports Qwen3.5-4B ahead on several listed metrics, including Global-MMLU PT in both modes.

The evaluations use different datasets and aggregation rules. The card documents their populations and limitations, including 100 selected questions for BrowseComp-Plus, 503 for LongBench v2, and 200 single-turn prompts plus 100 three-turn conversations for CAPITU. It says confirmed context-overflow cases on LongBench v2 count as zero and remain in the evaluation population. The card gives recommended sampling settings and notes that Think can use a higher output limit for harder problems. It does not state a commercial API price; the presented route is downloading and serving the weights.

## What the archive adds
New Runtime's earlier coverage has tracked open-weight releases for agent and coding workloads, including [MiniMax M3.1 Flash](https://newruntime.com/posts/minimax-releases-open-weights-for-minimax-m3-1-4125a5/) on 2026-10-23 and [Kolibri](https://newruntime.com/posts/kolibri-has-landed-a-sovereign-open-weight-model-55b53f/) on 2026-10-24. S5-nano is a new event, not a follow-up: its differentiator is Maritaca's native Brazilian Portuguese evaluation and locally deployable small-model formats.

## Trend signal
This does not yet fit a pattern New Runtime tracks.

## Why now
The model card provides enough serving detail to make local experimentation practical now: BF16, FP8 and GGUF files are listed alongside framework-specific launch instructions. That timing matters for teams evaluating whether a compact Portuguese model can meet workload needs without waiting for a hosted inference provider.

## Why it matters
For teams serving Brazilian Portuguese, S5-nano offers a testable alternative to routing every request to a general hosted model. Its two modes let builders compare deliberative answers with direct responses, while the quantized and GGUF variants create options for different hardware budgets and runtime preferences. The 262,144-token context ceiling may also suit document-heavy tasks, although maximum context is not a promise of useful quality or affordable throughput at that length.

The practical decision is not whether the aggregate benchmark table declares a winner. It is whether this checkpoint handles a team's own language, domain, latency and reliability requirements at an acceptable serving cost. Because the scores are published by the model maker, teams should reproduce a representative evaluation, include failure cases, and compare resource use across formats before changing production routing.

## Limits and uncertainty
The benchmark results are reported by Maritaca, and the model card does not establish independent replication. The published averages combine heterogeneous tasks; the card itself notes dataset-specific populations, grading rules and exclusions, so a single mean cannot predict performance for a particular product. The source gives no production latency, throughput, hardware-cost comparison, or commercial API price. It also does not provide evidence that any inference provider currently hosts S5-nano. Local deployment capability is documented, but operational performance will depend on hardware, context length, quantization and serving configuration.

## Skill to build now
Build a small Portuguese evaluation set from real user requests before choosing a mode or quantization. Run the same cases through Think and Instruct, record answer quality and failure types, then compare BF16 with an available lower-memory format on your target hardware. The model card documents both mode settings and serving examples, making that a concrete one-week test rather than a speculative migration.

## Reader takeaway
S5-nano is a practical open-weight release to evaluate if your product serves Brazilian Portuguese and you can run models locally. Maritaca's results are promising on its native-language benchmarks, but they are not independent proof of production quality. Test your workload, hardware and serving cost before treating the headline averages as a deployment decision.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models topic hub.
- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)