A 4B reasoning model that runs entirely on your machine. No account, no cloud, no telemetry — your conversations never leave the room they were typed in.

Why this model exists

Most people who ask an AI a serious question are asked to trust a server they cannot see. BrimUniverse Ai was built on the opposite premise: the machine that answers you should be the machine you own.

A model trained on high-quality public and custom datasets, designed to help you

What it does well

Deep reasoning. Built on a thinking-mode base, the model produces an explicit <think> chain before answering. You can read how it got there, not just where it landed. Naturally, it is best to use programs that display the answer directly, to avoid wasting time watching how the model "reasons."
Long, coherent conversations. Fine-tuned on multi-turn material where reasoning is carried across several exchanges, not restarted at every message.
Pragmatic. Trained to Trained in naturalness, but also in mathematics, science, and higher cognitive abilities.

Model details

Developed by: DoctorEdoP369
Base model: DavidAU/Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING
Architecture: Qwen3.5, 4B dense, thinking mode
Method: LoRA fine-tuning with Unsloth, trained on responses only
Quantization: Q6_K and Q8, from a bf16 merge
Also a setup for windows:
License: Apache 2.0

Training data

Fine-tuned on a large mixed corpus.

nvidia/Nemotron-Post-Training-Dataset-v2

Used to strengthen multi-turn reasoning and general answer quality.

License/Terms of Use: the dataset carries per-sample license information. It is predominantly CC-BY-4.0, with a small subset of prompts from WildChat under ODC-BY and a small subset from StackOverflow under CC-BY-SA.

This dataset contains synthetic data created using DeepSeek-R1-0528, Qwen2.5-14B-Instruct, Qwen2.5-32B-Instruct-AWQ, Qwen3-30B-A3B and Qwen3-235B-A22B. If this dataset is used to create, train, fine-tune, or otherwise improve an AI model, which is distributed or made available, such AI model may be subject to redistribution and use requirements in the Qwen License Agreement and the DeepSeek License Agreement.

Data Developer: NVIDIA

facebook/natural_reasoning

Used to broaden open-ended reasoning across domains.

Released by Meta under CC-BY-NC-4.0. This is a non-commercial license. Users who intend to deploy this model commercially should review the terms on the dataset page and assess their own position.

The remaining training material consists of custom high-quality datasets built for this project, including original Italian multi-turn reasoning data.

High-quality custom datasets.

Limitations and honest notes

This AI model is very powerful but this model has reduced safety filtering inherited from its base. It will answer questions that hosted assistants decline. That is a deliberate design choice aimed at autonomy — and it means the responsibility for what you ask, and what you do with the answer, is entirely yours.
It can be confidently wrong. The visible reasoning chain makes errors easier to spot, but it does not prevent them. The author does not take any responsibility for the use you will make of it.
It's not a professional. Nothing it says replaces a doctor, a lawyer, an accountant or an engineer.
Vision support depends on your client. Several local apps ignore the projector file entirely.

AI can make mistakes; it does not replace professional advice.

Created by DoctorEdoP369

Thanks to DavidAu and the Creators of Qwen3.5.

DoctorEdoP369/BrimUniverseAi

BrimUniverse Ai