Qarvexium

Who are we?

Qarvexium is an independent AI research and engineering organization focused on building efficient, accessible, and genuinely useful machine-learning systems — from language models and computer vision to speech, OCR, and generative AI.

We build things from the ground up, experiment aggressively, and release what we can for others to learn from, use, and improve.

What We Build

• reFLEX — our new family of modular neural architectures designed around objective understanding, selective experience retrieval, and adaptive information use.
• QED — our family of decoder-only language models, including base and instruction-tuned variants.
• Qanvas — lightweight generative vision research focused on making text-to-image systems practical on constrained hardware.
• QOCR — compact OCR models designed for efficient text recognition.
• Qlip / Q-VAE — supporting components for lightweight multimodal and generative systems.
• Experimental ML systems — architectures, tokenizers, datasets, training pipelines, and other research projects.

Our Best Architecture (reFLEX)

reFLEX is Qarvexium's current flagship architecture and our strongest exploration into building capable AI systems without relying solely on massive parameter counts.

reFLEX uses multiple specialized neural components that work together:

Main · Intent · Experience

• Main handles primary language modeling and generation.
• Intent focuses on identifying the objective behind the input.
• Experience retrieves potentially useful information related to that objective.
• Main decides whether the retrieved experience is relevant and worth using before generating a response.

Unlike a conventional monolithic language model, reFLEX gives different parts of the system different responsibilities while training them together as a unified architecture.

Our early experiments have been particularly promising at very small scales. A roughly 15M-parameter reFLEX model has demonstrated coherent responses and measurable improvements from its Experience component on a held-out evaluation set.

We're now exploring whether this architectural approach can continue to scale while maintaining the efficiency and compactness that make reFLEX interesting in the first place.

Release Date: September 3, 2026.

Examples from the 15M Model

PROMPT: A: Hey, what are you doing tonight?\nB:

GEN: " I'm going to watch some old movies from my childhood. I'm really excited about my weekend.\nA: That sounds fun! I wish I could go more often.\nB: I'm"

PROMPT: A: Where were you last night?\nB:

GEN: " I went to the mountains. I wanted to see some amazing fish!\nA: I've never been to the mountains. I've never been, but I'd love to go sometime.\nB"

The model can occasionally continue the conversation beyond the requested response, generating an additional A:/B: turn because the training format does not use an explicit end-of-response token. This is an inference-format limitation rather than the model completely failing to produce a relevant response, and can be handled by stopping generation when the next A: turn is detected.

Despite these limitations, we find the model's behavior particularly promising given its extremely small size. The current model is only around 15M parameters; for comparison, GPT-2 Small has approximately 124M parameters.

The goal of reFLEX is not simply to make smaller models. It is to explore whether architectural specialization can make limited parameters more effective.

Our Philosophy

We care about building, testing, and learning.

That means experimenting with new architectures, training models on limited hardware, optimizing inference, developing our own tokenization approaches, and sometimes throwing an idea away when the results prove it doesn't work.

Not every experiment becomes a polished product.

That's part of the point.

We believe useful AI research doesn't have to start with enormous budgets or enormous models. Small, carefully designed systems can still be interesting, capable, and worth exploring.

On Hugging Face

This organization is where we share our models, datasets, tokenizers, and other AI work.

Some repositories are experimental. Others are intended for practical use. Each project may have different goals, licenses, and levels of maturity, so please check the individual repository README and license before using anything.

Building in Public

Qarvexium is an evolving research ecosystem.

We're interested in:

Language Models · Computer Vision · OCR · Speech · Multimodal AI · Generative AI · Efficient ML · Tokenization · Model Architecture

Build it. Train it. Break it. Learn from it. Build something better.

---

Qarvexium — independent AI research & engineering.