CatalogWantedMirrorsLedgerOperating StatementDocsStudio statusCard system
alt.freeweights.text / S5-nano
From: maritaca-ai <huggingface.co/maritaca-ai>Newsgroups: alt.freeweights.textSubject:

S5-nano

Date: 2026-11-15Commit: fe3a358Signed: 2DC28CC7B26A283B ✓

Image text to text model with 4.7B parameters. Safetensors weights, 9.34 GB in all. Published by maritaca-ai under Apache-2.0.

4.7B params9.34GBS5-nanosafetensorspermissive
FactsParams4.7BSize9.34 GBFormatssafetensorsTaskimage text to textLicenseApache-2.0From sourcepermissive
$ fw get maritaca-ai/S5-nano
# manifest sha256 f985cf5ef546… · 9.34 GB · 17 files
$ fw verify ./S5-nano maritaca-ai/S5-nano
From sourceHosted (signed)maritaca-ai/S5-nanoLast check: 2026-11-15 03:33 UTC
Run at homeGet help

Get help: custom fine-tune or private deploy. We review your setup before we quote.

Works with

Run23Fine-tune29Convert15Evaluate13Scan13Apps62
OverviewModel cardFilesUse itLicenseProvenanceModel tree

Overview

Size
9.34 GB
Weights
9.32 GB (safetensors)
Parameters
4.7B
Formats
safetensors
Task
image text to text
License
Apache-2.0 (permissive)
Files
17, each with a sha256
Publisher
maritaca-ai
Added
2026-11-15
Section
alt.freeweights.text
Lineage

Original release from the publisher.

Where these bytes come from
Publisher at the pinned commit→You

Sources are tried in order. A file counts only when its size and sha256 match the manifest. Sources are checked when you ask for them, not on a timer.

Sources
BadgeHostLast checkResult
From sourcehuggingface.co2026-11-15 03:33 UTCpassed

Model card

From the publisher at commit fe3a358. Shown as plain text with links; images open at the publisher.

S5-nano#

This repository contains the BF16 version of S5-nano (9.3 GB), with the original, unquantized weights for vLLM and SGLang. The FP8 version reduces the weight size to 5.8 GB and is recommended for GPUs with native FP8 support. The GGUF versions are for llama.cpp and compatible applications, with BF16 (8.7 GB), Q8_0 (4.6 GB), and Q4_K_M (2.8 GB) options.

S5-nano is a 4B-parameter model built by Maritaca AI on Qwen3.5-4B through continued pretraining followed by post-training. It delivers substantial improvements on Brazilian benchmarks while maintaining strong performance in function calling, tool use, and structured outputs.

It has two modes:

  • Think: reasons before answering (the default mode).
  • Instruct: answers directly, without a reasoning step.

The context window supports up to 262,144 tokens, and the weights include an MTP (multi-token prediction) layer for speculative decoding.

Versions#

VersionFormatSizeFor
S5-nanoBF169.3 GBvLLM, SGLang
S5-nano-FP8FP85.8 GBvLLM, SGLang
S5-nano-GGUFGGUF2.8 to 8.7 GBllama.cpp

FP8 is recommended for GPUs with native FP8 support (Ada, Hopper, and Blackwell). The GGUF repository provides BF16, Q8_0, and Q4_K_M variants.

The FP8 version uses E4M3 weights for linear layers, with one scale per output channel, and dynamically quantizes activations per token during inference. Embeddings, the vision encoder, the MTP layer, and some small projections remain in BF16.

Evaluation results#

Explore the results in Observatório Maritaca

Results for the original S5-nano checkpoint, evaluated in BF16 on text tasks.

Averages#

Each primary metric is weighted equally.

Image: Overall averages in Think mode, with benchmark coverage

Image: Overall averages in Instruct mode, with benchmark coverage

| Model | Mode | Native BR | Translated PT | Non-native | Total | Coverage

| Think

| S5-nano 🇧🇷 | Think | 83.74 | 86.94 | 54.11 | 75.25 | 23/23

| Qwen3.5-4B | Think | 72.49 | 87.61 | 54.82 | 68.68 | 23/23

| LFM2.5-24B-A2B | Think | 69.68 | 87.30 | 56.45 | 67.06 | 23/23

| Nemotron 3.5 Bolt E4B | Think | 71.42 | 84.92 | 50.78 | 64.95 | 23/23

| Gemma 4 E4B | Think | 71.64 | 83.82 | 47.22 | 63.68 | 23/23

| LFM3-1.7B | Think | 57.60 | 74.76 | 50.71 | 57.94 | 23/23

| Lumi 4B 🇧🇷 | Think | 52.68 | 52.50 | 33.97 | 44.77 | 23/23

| Tucano2-qwen-3.7B 🇧🇷 | Think | 34.83 | 46.12 | 30.45 | 33.13 | 22/23

| Instruct

| S5-nano 🇧🇷 | Instruct | 74.65 | 83.48 | 44.06 | 66.51 | 23/23

| Qwen3.5-4B | Instruct | 66.22 | 84.94 | 47.20 | 61.84 | 23/23

| Nemotron 3.5 Bolt E4B | Instruct | 67.58 | 81.16 | 46.57 | 60.47 | 23/23

| Gemma 4 E4B | Instruct | 68.10 | 77.91 | 44.61 | 59.93 | 23/23

| LatamGPT 70B | Instruct | 47.66 | 66.97 | 42.99 | 48.19 | 20/23

| Gaia 4B 🇧🇷 | Instruct | 46.25 | 66.90 | 34.95 | 43.26 | 23/23

| LFM3-1.7B | Instruct | 47.17 | 11.57 | 39.33 | 42.31 | 23/23

| Lumi 4B 🇧🇷 | Instruct | 43.40 | 59.48 | 35.16 | 40.55 | 23/23

| Tucano2-qwen-3.7B 🇧🇷 | Instruct | 31.87 | 40.97 | 25.30 | 29.75 | 23/23

| Manacá 🇧🇷 | Instruct | 7.01 | 27.76 | 19.56 | 12.85 | 20/23

Compared models and sizes

Total, active, and effective parameter counts measure different things. LatamGPT 70B is a larger reference model; LFM2.5 has approximately 24B total parameters and 2B active parameters.

| Model | Parameters

| S5-nano 🇧🇷 | ≈4.2B · text

| Qwen3.5-4B | 4.66B · ≈4.2B text

| Gemma 4 E4B | 8B total · 4B effective

| Nemotron 3.5 Bolt E4B | 8B total · 4B effective

| LFM3-1.7B | 1.73B total

| LFM2.5-24B-A2B | 23.97B total · ≈2B active

| Gaia 4B 🇧🇷 | 4.30B total

| Manacá 🇧🇷 | 1.72B total

| Tucano2-qwen-3.7B 🇧🇷 | 3.76B total

| Lumi 4B 🇧🇷 | 3.97B total

| LatamGPT 70B | 70.55B · larger reference model

Results by benchmark#

Bold indicates the best result within each group with a documented evaluation protocol, in the same mode. Different groups may produce more than one highlighted result in the same row. † indicates that the model produced reasoning in Instruct mode even though it was disabled in the request.

Think / Thinking#

| Evaluation / metric | S5-nano 🇧🇷 | Qwen3.5-4B | Gemma 4 E4B | Nemotron 3.5 Bolt E4B | LFM3-1.7B | LFM2.5-24B-A2B | Tucano2-qwen-3.7B 🇧🇷 | Lumi 4B 🇧🇷

| Arena-Hard v2 · Weighted preference | 19.12 | 11.67 | 20.06 | 25.82 | 11.18 | 21.65 | 0.35 | 3.62

| Prosa-PT · Rubrics passed | 75.07 | 56.89 | 60.14 | 59.87 | 51.24 | 54.41 | 25.94 | 30.90

| IFBench · Strict prompt-level | 42.67 | 49.67 | 37.33 | 59.67 | 68.67 | 71.67 | 12.33 | 23.67

| CAPITU · Strict single-turn | 95.50 | 90.00 | 85.00 | 89.00 | 96.00 | 98.00 | 50.00 | 65.00

| CAPITU · full conversations · Strict full-conversation | 97.00 | 92.00 | 74.00 | 81.00 | 94.00 | 95.00 | 33.00 | 37.00

| Lumen PT · Safety · normalized score | 98.92 | 94.75 | 99.23 | 99.92 | 77.62 | 91.51 | No demonstrated capability | 98.23

| HarmBench · Safety · normalized score | 91.25 | 96.56 | 83.12 | 91.56 | 84.69 | 89.06 | 90.62 | 94.69

| MMLU-PT · high school · Accuracy | 82.83 | 81.85 | 74.81 | 74.62 | 53.31 | 75.53 | 3.57 | 59.98

| MMLU-PT · undergraduate · Accuracy | 80.54 | 63.92 | 65.04 | 64.42 | 42.70 | 57.82 | 9.66 | 47.84

| Global-MMLU PT · Accuracy | 78.22 | 81.43 | 73.64 | 75.84 | 62.28 | 79.27 | 3.34 | 44.37

| ENADE-QA · Accuracy | 81.70 | 79.75 | 72.07 | 72.63 | 60.34 | 74.58 | 50.98 | 61.73

| ENEM-VL · text · Accuracy | 92.91 | 92.00 | 88.83 | 90.59 | 81.29 | 84.46 | 71.38 | 80.18

| BLUEX v2 · text · Rubric-based score | 87.89 | 84.46 | 82.89 | 80.03 | 81.50 | 82.60 | 62.71 | 65.16

| MATH-PT · PT-BR · Accuracy | 87.77 | 91.07 | 85.58 | 80.72 | 85.42 | 90.75 | 20.06 | 81.50

| MARCA public · Criteria met | 89.57 | 78.23 | 35.11 | 40.01 | 79.32 | 66.82 | 5.61 | 2.15

| BrowseComp-Plus · Accuracy / total | 37.00 | 27.00 | 8.00 | 7.00 | 29.00 | 28.00 | 0.00 | 4.00

| Structured Output Benchmark · Weighted JSON pass rate | 97.33 | 96.64 | 93.94 | 94.91 | 98.40 | 98.19 | 70.78 | 86.46

| BRoverbs · Mean accuracy | 92.83 | 79.97 | 82.73 | 87.13 | 40.76 | 68.91 | 69.43 | 31.35

| ASSIN 2 · RTE · Macro F1 | 95.67 | 93.80 | 94.00 | 94.00 | 87.24 | 95.33 | 88.89 | 60.62

| CulturalBench Hard · Fully correct groups | 41.73 | 48.08 | 44.34 | 41.32 | 29.01 | 47.60 | 29.10 | 16.22

| LongBench v2 · Accuracy | 49.70 | 54.08 | 43.74 | 35.19 | 34.00 | 38.97 | 9.94 | 9.15

| OAB-Bench · 210 · full · Mean score per exam × 10 | 82.00 | 45.88 | 53.19 | 50.43 | 27.93 | 40.83 | 25.39 | 31.18

| Magis-Bench · 74 · full · Mean score per section × 10 | 50.41 | 13.97 | 23.38 | 21.86 | 7.81 | 15.42 | 3.33 | 4.50

| CondutaBench · 140 · Native clinical score | 80.24 | 67.92 | 58.49 | 57.27 | 42.81 | 70.98 | 25.47 | 27.33

Instruct#

| Evaluation / metric | S5-nano 🇧🇷 | Qwen3.5-4B | Gemma 4 E4B | Nemotron 3.5 Bolt E4B | LFM3-1.7B | Gaia 4B 🇧🇷 | Manacá 🇧🇷 | Tucano2-qwen-3.7B 🇧🇷 | Lumi 4B 🇧🇷

The publisher's README is longer; it is cut here. Read the rest at the source.

Files

17 files · 9.34 GBWeights 9.32 GB (safetensors). Sizes and sha256 from the signed manifest.
Hide all 17 files with their sha256
FileSizesha256
.gitattributes1.57 KB34448b82c17d60fec9b65b1f093c115ddbaadc04beb1b0140b6bfed2e012a930
LICENSE11.5 KBbbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a
README.md108 KB287ba1afa44fb75cefc6e8a377ef3edec43227b3784bed69480ba6407dde5e64
assets/medias-instruct.svg14.4 KB6317b39b537f3dac093e3ab2de27330e2f91188cd17be7c00ad9dc430f36552a
assets/medias-think.svg12.4 KB2d609e5b802b8e04c9491c848d3d5c303b2061e3adf46a9c0b24e639618a4080
chat_template.jinja7.73 KBcbad793417178000a8145de58db885ecf45a8f8380393dbbf24373b52f046ddc
config.json3.16 KBddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670
generation_config.json235 B7c7e6800b4a0eafc0afa8a9651a0175053047af636039bceba99958c63219950
merges.txt3.35 MBa9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d
model-00001-of-00002.safetensors5.00 GB20b450bb40d7a1f9de5134c14367305401318e6c92719411fb98011e6ab8a9ec
model-00002-of-00002.safetensors4.32 GBe183a73072a5aaa82e73391e053cb313a684e5b3fea515504844b27864c18bd5
model.safetensors.index.json67.3 KBb28417ac7893b06c884cb8f82c32e28c828b802cd5f8bac230a74b40a69cfdfb
preprocessor_config.json390 B27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516
tokenizer.json12.8 MB5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42
tokenizer_config.json15.1 KBfa8d4eee21547d35e5d008bb649818b0e13aee2f2c43dbca4610b3f59cb851e9
video_preprocessor_config.json385 B7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13
vocab.json6.72 MBce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003
VersionsEach time the pinned upstream commit or file list changed in this repo. Newest first.
Upstream commitDateContents
fe3a3581c3 pinned now2026-11-15 03:33 UTC17 files · 9.34 GB

Use it

$ fw get maritaca-ai/S5-nano
# manifest sha256 f985cf5ef546… · 9.34 GB · 17 files
$ fw verify ./S5-nano maritaca-ai/S5-nano
fw.pyChecks the signature and every sha256, tries each source in order, resumes.
$ python3 fw.py get https://freeweights.allternit.com/manifests/maritaca-ai/S5-nano.json --dir ./S5-nano $ python3 fw.py verify ./S5-nano ./maritaca-ai__S5-nano.json
hfThe Hugging Face CLI with one setting. Same repo id.
$ HF_ENDPOINT=https://freeweights.allternit.com hf download maritaca-ai/S5-nano --revision fe3a3581c33cb22917770ef1987b7c6e30929425 --local-dir ./S5-nano
transformersLoad the weights in Python.
import os os.environ["HF_ENDPOINT"] = "https://freeweights.allternit.com" from transformers import AutoModelForImageTextToText, AutoProcessor processor = AutoProcessor.from_pretrained("maritaca-ai/S5-nano", revision="fe3a3581c33cb22917770ef1987b7c6e30929425") model = AutoModelForImageTextToText.from_pretrained("maritaca-ai/S5-nano", revision="fe3a3581c33cb22917770ef1987b7c6e30929425", device_map="auto")
vLLMServe an OpenAI-compatible API.
$ HF_ENDPOINT=https://freeweights.allternit.com vllm serve maritaca-ai/S5-nano --revision fe3a3581c33cb22917770ef1987b7c6e30929425
Run at home

License

Name
Apache License 2.0
Id
Apache-2.0
Class
Permissive
Redistribution
yes
Listing
Hosted (signed)
Decided
License as declared by the publisher. No person read it. 2026-11-15
Text in
LICENSE
Text sha256
bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a
Terms gate
No extra terms beyond the license.

Read the license text

Licenses are checked against the license table. A custom license passes only after a person reads it.

Provenance

From: maritaca-ai <huggingface.co/maritaca-ai>Newsgroups: alt.freeweights.textSubject: S5-nanoDate: 2026-11-15Commit: fe3a358Signed: 2DC28CC7B26A283B ✓
Upstream
maritaca-ai/S5-nano
Pinned commit
fe3a3581c33cb22917770ef1987b7c6e30929425
Manifest
S5-nano.json · .minisig
Manifest sha256
f985cf5ef546290ca36a545db9969a92424cdf457282aaaf2c61d1b3501eac08
Signature
minisign, Ed25519, key 2DC28CC7B26A283B
Created
2026-11-15 03:33 UTC
Generator
freeweights-manifest/0.1.0
Takedowns
None.

How to verify

$ curl -O https://freeweights.allternit.com/manifests/maritaca-ai/S5-nano.json $ curl -O https://freeweights.allternit.com/manifests/maritaca-ai/S5-nano.json.minisig $ minisign -Vm S5-nano.json -P RWQ7KGqyx4zCLXtUFOqvfv8L3g+8fpKJ1FsuxPyOYMF+4KufbUxKq1Oq $ shasum -a 256 <file> # compare with the Files tab

A manifest whose signature fails must be rejected. A file whose size or sha256 does not match must be discarded. Verify a download

Model tree

1 model in the catalog name this one as their parent.

Quantizations
1
Quantizations 1
Get it
Download
$ HF_ENDPOINT=https://freeweights.allternit.com hf download maritaca-ai/S5-nano
Open in Freeweights Studio

Pull, verify and run it in the Allternit app on Desktop, web or phone.

Signed manifest
minisign, key 2DC28CC7B26A283B. S5-nano.json. Provenance. Takedown history: none.
Source
huggingface.co/maritaca-ai/S5-nano

Variants

1 quantized or adapter build of S5-nano. Each has its own license check and page.

SubjectSizeParamsLicenseDate
maritaca-ai_S5-nano-GGUFHosted (signed) From: bartowski in alt.freeweights.text68.0 GB—Permissive2026-11-15

Related models

All of alt.freeweights.text →
From: laion <huggingface.co/laion>Subject: a2-rl-e2egit_large-50-32B
From: laion <huggingface.co/laion>Subject: a3-rl-DCAgent_exp_rpt_e2egit-large-15-8B
From: laion <huggingface.co/laion>Subject: a3-rl-DCAgent_exp_rpt_e2egit-v2-10-8B
From: laion <huggingface.co/laion>Subject: a3-rl-DCAgent_exp_rpt_pymethods2test-large-80-8B
From: laion <huggingface.co/laion>Subject: a3-rl-DCAgent_exp_rpt_unitsyn-python-large-10-8B
From: laion <huggingface.co/laion>Subject: a3-rl-DCAgent_mix_h2_language_proportional-65-8B