Overview
Lumi-4B is a 4B-parameter instruct language model developed by CEIA/UFG’s NLP team, with a focus on Brazilian Portuguese and support for English. It starts from Nemotron 3 Nano 4B and was trained with continual pre-training, knowledge distillation, and supervised fine-tuning. Maintainer: ceia-research.
Best use cases
Conversational and instruction-following text generation, especially in Brazilian Portuguese; English text generation is also an intended use. The training pipeline was designed to retain mathematics and coding capabilities, but those tasks should be evaluated for the intended application.
Limitations
Keep thinking disabled for this release. It has no additional RL/RLVR or dedicated model-identity training; tool use is not documented as validated. The model may produce inaccurate, biased, or unsupported content, and performance outside Brazilian Portuguese and in specialized domains needs evaluation. The README describes a later v0.2 release separately; its updated data blend and benchmark claims should not be assumed to apply to this exact model.
How it compares
The supplied related-model list includes GAIA Gemma-3 Gaia PT-BR 4B, described as a Brazilian-Portuguese model based on Gemma 3, and Manacá-1B base, described as a Brazilian-Portuguese base model. These descriptions establish differences in stated model focus and type, not comparative quality or capability.
Technical specifications
4B instruct model; languages: Portuguese and English. The stated starting point is Nemotron 3 Nano 4B, with Nemotron 3.5 Lightning used as the distillation teacher. CPT used 50 billion tokens and an 8,192-token training context, followed by SFT. Architecture and tokenizer details are referred to the base model documentation but are not specified here. No hardware, validated inference versions, recommended generation settings, or measured benchmark values are provided in text.
Model inputs and outputs
Inputs
Chat messages using the released checkpoint’s chat template. The supplied vLLM example uses an OpenAI-compatible chat-completions request. Set enable_thinking to false; the example’s other generation settings are illustrative.
Outputs
Generated text in response to chat or instruction prompts. The model is tagged for text generation; no structured-output format or other output controls are documented.
Getting started
The README provides a vLLM template, but explicitly says to validate the repository ID, runtime version, and chat template against the released checkpoint. Its repository ID is presented as a placeholder, so no reliable runnable setup example is supplied.
Frequently asked questions
Q: What is Lumi-4B intended for?
A: It is an instruct model for conversational and instruction-following text generation, focused on Brazilian Portuguese and also intended for English.
Q: Should thinking be enabled?
A: No. Thinking must remain disabled for this release.
Q: What training was used?
A: The model underwent continual pre-training with knowledge distillation, followed by supervised fine-tuning. The described CPT used 50 billion tokens.
Q: Does this release include reinforcement learning?
A: No additional RL or RLVR is included in this release. Such training is listed as a planned next step.
Q: Is tool use validated?
A: The supplied model information says tool-use training and evaluation are undocumented, so treat it as unvalidated.
Q: What context length can I use at inference?
A: The training context is stated as 8,192 tokens, but a validated inference context length is not provided.
Q: Are benchmark results available?
A: The README references benchmark images, but the supplied text gives no numerical results. It also describes v0.2 separately, so do not attribute that version’s claims to this model.
This is a simplified guide to an AI model called lumi-4b maintained by ceia-research. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter.
