github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 08:30:05 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #14533] BERTimbau, Sabiá #9428

New Issue
Closed
opened 2026-04-12 22:21:16 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-12 22:21:16 -05:00
Owner
Copy Link

Originally created by @guicybercode on GitHub (Mar 1, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14533

[Feature Request] Add Brazilian Portuguese (PT-BR) native LLM models to the Ollama Library

Summary

The Ollama official library currently has zero natively trained Brazilian Portuguese models. With 260M+ Portuguese speakers worldwide (215M+ in Brazil alone) and a growing ecosystem of open-source PT-BR LLMs, adding these models would significantly expand Ollama's reach in Latin America's largest AI market.

The Problem

While models like Llama 3.2/3.3/4, Gemma, and Qwen include Portuguese as a "supported language," they are not optimized for Brazilian Portuguese. This leads to:

  • Poor tokenizer efficiency: Models using English-centric tokenizers (e.g., Llama 2's BPE) encode Portuguese text very inefficiently — often 2-3x more tokens than necessary for the same content (research from Tucano paper).
  • Cultural blindspots: Mainstream models struggle with Brazilian-specific concepts (PIX payment system, CPF/CNPJ, ENEM, vestibular, regionalismos, gírias).
  • Lack of local compliance awareness: Brazilian enterprises need models that understand LGPD (Brazil's data protection law) and local regulatory frameworks.

Currently, the only PT-BR options on Ollama are community-uploaded models (e.g., splitpierre/bode-alpaca-pt-br, brunoconterato/Gemma-3-Gaia-PT-BR-4b-it, RecognaNLP/chatbode) — none are in the official ollama.com/library.

Proposed Models for Inclusion

Tier 1 — Strong candidates (peer-reviewed, Apache 2.0, active development)

Model Params License HuggingFace Notes
Tucano (160m, 630m, 1.1B, 2.4B) 160M–2.4B Apache 2.0 TucanoBR Natively pre-trained on GigaVerbo (200B tokens PT-BR). Published in Patterns (Cell Press). Includes instruct variants and ViTucano (vision).
TeenyTinyLlama 160M, 460M Apache 2.0 nicholasKluge First open-source tiny LLMs natively trained in PT-BR. Same team as Tucano.
GAIA 4B (Gemma 3 base) Open Google DeepMind / CEIA-UFG Google DeepMind + Universidade Federal de Goiás. Already piloted by Brazilian government (TCM-GO) and Unimed.

Tier 2 — Established models (Llama-based fine-tunes)

Model Params License HuggingFace Notes
Canarim (7B, 7B-Instruct) 7B Llama 2 Community dominguesm Extended training on 16B tokens from Portuguese Common Crawl. Vestibulaide variant for Brazilian entrance exams.
Cabrita 3B Apache 2.0 22h/cabrita-lora-v0-1 Fine-tuned OpenLLaMA 3B with custom PT-BR tokenizer.
Bode 7B, 13B Llama 2 Community recogna-nlp LoRA fine-tune of Llama 2 on translated Alpaca dataset.

Tier 3 — Notable but limited availability

Model Notes
Sabiá (7B, Sabiá-2, Sabiá-3) By maritaca.ai (Brazilian AI startup). Sabiá-3 is commercial API only. Sabiá-7B is Llama-based and open.
Gervásio (7B) European + Brazilian Portuguese variants. Llama 2 derivative.
Mula Sparse MoE architecture, natively trained in PT-BR. Early stage.

Why This Matters

  1. Market size: Brazil is the 5th largest country by population and the largest economy in Latin America. Portuguese is the 6th most spoken language globally.
  2. Growing AI ecosystem: Brazil's national AI plan (2024–2028) is actively investing in local AI development. The state of Goiás became the first Brazilian state to create an AI regulatory framework.
  3. Developer demand: The Open-PT-LLM-Leaderboard on HuggingFace has 100+ entries specializing in Portuguese, showing massive community interest.
  4. Edge/local deployment: Brazilian enterprises (banks, healthcare, government) increasingly need local LLM deployment for LGPD compliance — exactly Ollama's sweet spot.
  5. Tokenizer efficiency: Natively trained models like Tucano use optimized PT-BR tokenizers, meaning faster inference and lower memory usage for Portuguese text compared to multilingual models.

Suggested Starting Point

The Tucano series would be the ideal first addition:

  • ✅ Apache 2.0 license
  • ✅ Peer-reviewed publication in Patterns (Cell Press, July 2025)
  • ✅ Multiple sizes (160M to 2.4B) — perfect for edge deployment
  • ✅ Instruct variants available (Tucano-1b1-Instruct, Tucano-2b4-Instruct)
  • ✅ Vision variant (ViTucano)
  • ✅ Full open-source training code + evaluation harness on GitHub
  • ✅ Custom Portuguese tokenizer with superior compression
  • ✅ Already outperforms Llama-3.2-1B on Portuguese benchmarks

References

  • Corrêa et al. (2025). "Tucano: Advancing Neural Text Generation for Portuguese." Patterns, Cell Press. arXiv:2411.07854
  • Corrêa et al. (2024). "TeenyTinyLlama: open-source tiny language models trained in Brazilian Portuguese." SoftwareX. DOI:10.1016/j.softx.2024.101724
  • Assis et al. (2025). "Exploring Brazil's LLM Fauna." Journal of the Brazilian Computer Society, 31(1). DOI:10.5753/jbcs.2025.5814
  • GAIA announcement: Google DeepMind + CEIA-UFG collaboration (June 2025)

Labels: feature, models, multilingual

I'm happy to help with GGUF conversions, Modelfile creation, or testing if the team is interested in moving forward with any of these models.

Originally created by @guicybercode on GitHub (Mar 1, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14533 # [Feature Request] Add Brazilian Portuguese (PT-BR) native LLM models to the Ollama Library ## Summary The Ollama official library currently has **zero natively trained Brazilian Portuguese models**. With 260M+ Portuguese speakers worldwide (215M+ in Brazil alone) and a growing ecosystem of open-source PT-BR LLMs, adding these models would significantly expand Ollama's reach in Latin America's largest AI market. ## The Problem While models like Llama 3.2/3.3/4, Gemma, and Qwen include Portuguese as a "supported language," they are **not optimized** for Brazilian Portuguese. This leads to: - **Poor tokenizer efficiency**: Models using English-centric tokenizers (e.g., Llama 2's BPE) encode Portuguese text very inefficiently — often 2-3x more tokens than necessary for the same content ([research from Tucano paper](https://arxiv.org/html/2411.07854v1)). - **Cultural blindspots**: Mainstream models struggle with Brazilian-specific concepts (PIX payment system, CPF/CNPJ, ENEM, vestibular, regionalismos, gírias). - **Lack of local compliance awareness**: Brazilian enterprises need models that understand LGPD (Brazil's data protection law) and local regulatory frameworks. Currently, the only PT-BR options on Ollama are **community-uploaded models** (e.g., `splitpierre/bode-alpaca-pt-br`, `brunoconterato/Gemma-3-Gaia-PT-BR-4b-it`, `RecognaNLP/chatbode`) — none are in the official `ollama.com/library`. ## Proposed Models for Inclusion ### Tier 1 — Strong candidates (peer-reviewed, Apache 2.0, active development) | Model | Params | License | HuggingFace | Notes | |-------|--------|---------|-------------|-------| | **Tucano** (160m, 630m, 1.1B, 2.4B) | 160M–2.4B | Apache 2.0 | [TucanoBR](https://huggingface.co/TucanoBR) | Natively pre-trained on GigaVerbo (200B tokens PT-BR). Published in *Patterns* (Cell Press). Includes instruct variants and ViTucano (vision). | | **TeenyTinyLlama** | 160M, 460M | Apache 2.0 | [nicholasKluge](https://huggingface.co/nicholasKluge) | First open-source tiny LLMs natively trained in PT-BR. Same team as Tucano. | | **GAIA** | 4B (Gemma 3 base) | Open | [Google DeepMind / CEIA-UFG](https://huggingface.co/) | Google DeepMind + Universidade Federal de Goiás. Already piloted by Brazilian government (TCM-GO) and Unimed. | ### Tier 2 — Established models (Llama-based fine-tunes) | Model | Params | License | HuggingFace | Notes | |-------|--------|---------|-------------|-------| | **Canarim** (7B, 7B-Instruct) | 7B | Llama 2 Community | [dominguesm](https://huggingface.co/dominguesm/canarim-7b) | Extended training on 16B tokens from Portuguese Common Crawl. Vestibulaide variant for Brazilian entrance exams. | | **Cabrita** | 3B | Apache 2.0 | [22h/cabrita-lora-v0-1](https://huggingface.co/22h) | Fine-tuned OpenLLaMA 3B with custom PT-BR tokenizer. | | **Bode** | 7B, 13B | Llama 2 Community | [recogna-nlp](https://huggingface.co/recogna-nlp) | LoRA fine-tune of Llama 2 on translated Alpaca dataset. | ### Tier 3 — Notable but limited availability | Model | Notes | |-------|-------| | **Sabiá** (7B, Sabiá-2, Sabiá-3) | By maritaca.ai (Brazilian AI startup). Sabiá-3 is commercial API only. Sabiá-7B is Llama-based and open. | | **Gervásio** (7B) | European + Brazilian Portuguese variants. Llama 2 derivative. | | **Mula** | Sparse MoE architecture, natively trained in PT-BR. Early stage. | ## Why This Matters 1. **Market size**: Brazil is the 5th largest country by population and the largest economy in Latin America. Portuguese is the 6th most spoken language globally. 2. **Growing AI ecosystem**: Brazil's national AI plan (2024–2028) is actively investing in local AI development. The state of Goiás became the first Brazilian state to create an AI regulatory framework. 3. **Developer demand**: The [Open-PT-LLM-Leaderboard](https://huggingface.co/spaces) on HuggingFace has 100+ entries specializing in Portuguese, showing massive community interest. 4. **Edge/local deployment**: Brazilian enterprises (banks, healthcare, government) increasingly need local LLM deployment for LGPD compliance — exactly Ollama's sweet spot. 5. **Tokenizer efficiency**: Natively trained models like Tucano use optimized PT-BR tokenizers, meaning faster inference and lower memory usage for Portuguese text compared to multilingual models. ## Suggested Starting Point The **Tucano series** would be the ideal first addition: - ✅ Apache 2.0 license - ✅ Peer-reviewed publication in *Patterns* (Cell Press, July 2025) - ✅ Multiple sizes (160M to 2.4B) — perfect for edge deployment - ✅ Instruct variants available (Tucano-1b1-Instruct, Tucano-2b4-Instruct) - ✅ Vision variant (ViTucano) - ✅ Full open-source training code + evaluation harness on [GitHub](https://github.com/Nkluge-correa/Tucano) - ✅ Custom Portuguese tokenizer with superior compression - ✅ Already outperforms Llama-3.2-1B on Portuguese benchmarks ## References - Corrêa et al. (2025). "Tucano: Advancing Neural Text Generation for Portuguese." *Patterns*, Cell Press. [arXiv:2411.07854](https://arxiv.org/abs/2411.07854) - Corrêa et al. (2024). "TeenyTinyLlama: open-source tiny language models trained in Brazilian Portuguese." *SoftwareX*. [DOI:10.1016/j.softx.2024.101724](https://doi.org/10.1016/j.softx.2024.101724) - Assis et al. (2025). "Exploring Brazil's LLM Fauna." *Journal of the Brazilian Computer Society*, 31(1). [DOI:10.5753/jbcs.2025.5814](https://doi.org/10.5753/jbcs.2025.5814) - GAIA announcement: Google DeepMind + CEIA-UFG collaboration (June 2025) --- **Labels**: `feature`, `models`, `multilingual` I'm happy to help with GGUF conversions, Modelfile creation, or testing if the team is interested in moving forward with any of these models.
GiteaMirror commented 2026-04-12 22:21:18 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 1, 2026):

The Tucano series is based on llama2 architecture and is theoretically easy to import. However, both ollama and llama.cpp imports create models that generate poor output.

ollama import:

$ ollama run frob/tucano:2b4-fp16 hello
h[UNK_BYTE_0xe29681▁você]voc[UNK_BYTE_0xe29681▁me]mesA[UNK_BYTE_0xe29681▁Você]Voc
[UNK_BYTE_0xe29681▁o]o[UNK_BYTE_0xe29681▁o]o[UNK_BYTE_0xe29681▁eu]eu
[UNK_BYTE_0xe29681▁Eu]Eu.[UNK_BYTE_0xe29681▁Este]Esteart[h[UNK_BYTE_0xe29681▁você]voc
[UNK_BYTE_0xe29681▁me]mesA[UNK_BYTE_0xe29681▁Você]Voc[UNK_BYTE_0xe29681▁o]o
[UNK_BYTE_0xe29681▁o]

This looks like umapped codepoints in the tokenizer.

llama.cpp import:

$ ollama run frob/tucano:2b4-fp16 hello
h "helsin O aada euEu Eu eu A - você "Eu gostaria A é a é para O.: Resposta O que O artigoUm
o'Reseu: EuVocêQ "Oeremos</>:<pad> tem:</Resposta: "A: Analcreva Você é
</No.Resposta:"Res: Escreva um teste: CenaEu estava planejTipo: "Oi, eu eu tenho "Eu quero
você pode me encontre esta pergunta: Quem sãouitasarei:O que pode produzir ser um bom
amigo de um amigo? Por favor, responda à seguinte pergunta: "O que os alunos fazem em
uma sala? Uma pessoa quer fazer um bolo de chocolate para sua festa de aniversário Para
cada 2 xícaras de chocolate em um bolo, quanto açúcar deve ser adicionado?
</instruction: Você está preparando uma festa de aniversário e deseja fazer um bolo de
chocolate com 2 xícaras de chocolate e 1/2 xícara de açúcar, mas não quer usar 2 xícaras
de chocolate. Em seguida, você adiciona mais açúcar do que o necessário para a receita.
Quantas colheres de açúcar você precisará adicionar?

I'm not familiar with Portuguese but the existence of </ sequences leads me to believe
that this is not coherent.

<!-- gh-comment-id:3981013001 --> @rick-github commented on GitHub (Mar 1, 2026): The Tucano series is based on llama2 architecture and is theoretically easy to [import](https://github.com/ollama/ollama/blob/main/docs/import.mdx#importing-a-model-from-safetensors-weights). However, both ollama and llama.cpp imports create models that generate poor output. ollama import: ``` $ ollama run frob/tucano:2b4-fp16 hello h[UNK_BYTE_0xe29681▁você]voc[UNK_BYTE_0xe29681▁me]mesA[UNK_BYTE_0xe29681▁Você]Voc [UNK_BYTE_0xe29681▁o]o[UNK_BYTE_0xe29681▁o]o[UNK_BYTE_0xe29681▁eu]eu [UNK_BYTE_0xe29681▁Eu]Eu.[UNK_BYTE_0xe29681▁Este]Esteart[h[UNK_BYTE_0xe29681▁você]voc [UNK_BYTE_0xe29681▁me]mesA[UNK_BYTE_0xe29681▁Você]Voc[UNK_BYTE_0xe29681▁o]o [UNK_BYTE_0xe29681▁o] ``` This looks like umapped codepoints in the tokenizer. llama.cpp import: ``` $ ollama run frob/tucano:2b4-fp16 hello h "helsin O aada euEu Eu eu A - você "Eu gostaria A é a é para O.: Resposta O que O artigoUm o'Reseu: EuVocêQ "Oeremos</>:<pad> tem:</Resposta: "A: Analcreva Você é </No.Resposta:"Res: Escreva um teste: CenaEu estava planejTipo: "Oi, eu eu tenho "Eu quero você pode me encontre esta pergunta: Quem sãouitasarei:O que pode produzir ser um bom amigo de um amigo? Por favor, responda à seguinte pergunta: "O que os alunos fazem em uma sala? Uma pessoa quer fazer um bolo de chocolate para sua festa de aniversário Para cada 2 xícaras de chocolate em um bolo, quanto açúcar deve ser adicionado? </instruction: Você está preparando uma festa de aniversário e deseja fazer um bolo de chocolate com 2 xícaras de chocolate e 1/2 xícara de açúcar, mas não quer usar 2 xícaras de chocolate. Em seguida, você adiciona mais açúcar do que o necessário para a receita. Quantas colheres de açúcar você precisará adicionar? ``` I'm not familiar with Portuguese but the existence of `</` sequences leads me to believe that this is not coherent.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#9428
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?