Skip to content

Most AI in Brazil is an application layer on top of foreign models. Elvisia builds the layer underneath.

Tokenizer, Portuguese language models, router, agents and edge intelligence. Our own stack, built in Pernambuco.

Elvisia pt_completo_v48000
GPT-4 (cl100k)
Llama 3

The stack

Six layers, bottom up. Each one with where it actually stands today.

  1. PT-BR tokenizer

    48,000-token unigram with byte fallback, trained from scratch across nine domains. Native tokens for tool calling, fill-in-the-middle and source citation.

    Available

  2. Our own models

    NeuralBR-500 and NBR-1B are published. A ~2B model is training now on the new tokenizer.

    Available In development

  3. Model router

    Decides, per task, between a local Elvisia model, a specialised one, or an external frontier model.

    Research

  4. Agent runtime

    Goal, plan, tools, execution and verification. You delegate an outcome, not a conversation.

    In development

  5. Brazilian tool layer

    Company registries, official gazettes, public tenders, legislation, postcodes. The sources an agent needs to work in Brazil.

    Research

  6. Deployment

    Cloud, on-premise, edge on Jetson and Raspberry Pi, and genuinely offline.

    In development

The tokenizer

A 48,000-token unigram trained from scratch on a nine-domain mixture, with legal text as the largest slice. Byte fallback, so no text goes unrepresented.

Fertility measured by Elvisia, in tokens per word. Lower is better.
legal Portuguese1.08 tokens/word
general Portuguese1.17 tokens/word
medical Portuguese1.27 tokens/word
northeastern Portuguese1.28 tokens/word
English1.51 tokens/word
code2.15 tokens/word

Tokens it has that a general-purpose tokenizer does not:

  • <|tool_call|>
  • <|tool_result|>
  • <|fim_prefix|>
  • <|fim_middle|>
  • <|fim_suffix|>
  • <|cite|>
  • <|doc_start|>
  • <|doc_end|>

Twenty-five special tokens, designed for agents: tool call, tool result, fill-in-the-middle and source citation.

The models

Trained from scratch, not fine-tuned from a foreign model.

NeuralBR-500

Available
  • ~468M parameters, decoder-only
  • 24 layers, GQA with 16 heads and 8 KV heads
  • 2,048-token context
  • ~3.89M document corpus, ~1.5B tokens
  • pretrain, SFT, LoRA, DPO and GGUF
  • full training run for ~US$ 130 on H200

NBR-1B

Available
  • ~1 to 1.5B parameters
  • trained from scratch in Brazilian Portuguese
  • open weights on Hugging Face
  • multiple-choice and chat variants
Weights on Hugging Face

~2B

In development
  • training since September 2026
  • first model on the pt_completo_v48000 tokenizer
  • no publishable results yet

Where the stack runs

Drones and robots are not a separate product. They are one more place the same stack runs.

  • Cloud
  • On-premise
  • Edge (Jetson, Raspberry Pi)
  • Offline, no network at all
  • Embedded devices
  • Drones and robots, as an experimental platform

Photo of the test bench pending. No AI-generated image stands in for a real photo of the lab.

Elvisia App

In development

Research the web, check the sources and turn research into editable presentations.

Elvisia finds it for you. And does it for you.

  1. goal You say what you need, not how to do it.
  2. sniffing The agent plans, picks tools and searches the sources.
  3. found The answer shows the sources consulted.
  4. fetched The result includes visible sources so you can check what matters.
Create account Sign in

Beta in early access.

Elvisia can make mistakes. Check the sources before making important decisions.

Who it is for

Where privacy, offline operation and real Portuguese change the outcome.

Industry
The model runs on the factory floor, offline. Process data never leaves the building.
Healthcare
Local inference, without sending patient records to a third-party server abroad.
Government
An auditable stack, hosted where the law requires, with no single-vendor dependency.
Legal
The tokenizer was trained with legal text as the largest slice of the corpus: 1.08 tokens per word.
Education
Models small enough to run on the equipment schools already own.
Research
Open weights, a described method and measured numbers, for anyone who wants to reproduce them.

The lab

Elvisia is a Brazilian AI company and lab, based in Pernambuco, founded in 2026.

Open research lines are interpretability, alternative architectures, world models and embodied intelligence. Weights that are ready stay open on Hugging Face.

Contact

elvisia@elvisia.com.br Models on Hugging Face