VectraYX-Vision-1B: A Sub-2B Spanish/LATAM Cybersecurity Vision-Language Model with Structured Visual Reasoning and Native Tool Use

Imagine a world where computers are like incredibly smart librarians who have read every book ever written but have never actually seen a picture, a chart, or a screen. For a long time, these "text-only" librarians could answer questions about what a picture might contain based on descriptions, but they couldn't actually see the image itself. Then, a new type of computer brain called a "Vision-Language Model" (VLM) arrived. Think of these as librarians who finally got their eyes opened; they can look at a photo and read the text inside it, connecting what they see with what they know. This is a huge deal for fields like cybersecurity, where experts spend their days staring at complex screens filled with code, network maps, and security alerts. However, most of these "super-eyes" are giant, expensive, and only speak English, making them useless for smaller teams or those who need to keep their data completely private and offline.

This paper introduces a new, tiny, and specialized computer brain named VectraYX-Vision-1B. It's designed specifically for Spanish-speaking cybersecurity experts who need to analyze security screens without sending their data to the cloud. The researcher built this model to be small enough to run on a regular laptop (under 2 billion parameters) while still being able to "think" out loud about what it sees and even call for help from external tools. But here is the twist: the author is being incredibly honest about a major hiccup. While the model is brilliant at speaking Spanish and following instructions, it currently struggles to actually understand the pictures it's looking at. It's like a student who can write a perfect essay in Spanish but keeps staring at a blank page when asked to describe a diagram. The paper doesn't claim to have fixed this yet; instead, it releases the blueprint of the student, the test questions, and a specific mystery about how the student's brain is wired, inviting others to help solve the puzzle.

The Story of the Tiny, Spanish-Speaking Security Eye

The Problem: Blind Spots in Security
Security analysts are like digital detectives. They spend their days looking at screens filled with code, network traffic, and error logs. Sometimes they need to figure out if a piece of software is a virus, or if a network scan shows a hidden danger. Until now, the AI tools helping them were either too big to run on a local computer (requiring a connection to the cloud, which is a no-go for secret data) or they couldn't speak Spanish well. The existing "super-eyes" were also too general; they were great at describing a cat or a sunset but terrible at reading a complex security tool screen.

The Solution: A Custom-Built Detective
The author built VectraYX-Vision-1B, a model designed from the ground up for this specific job.

• It's Tiny: At just 1.04 billion parameters, it's small enough to fit on a regular computer, allowing it to work in "air-gapped" environments (computers that are physically disconnected from the internet for security).
• It Speaks Spanish: It was trained specifically to understand and speak Spanish, making it accessible for Latin American security teams.
• It Thinks Aloud: Before giving an answer, the model uses special tokens (like <|think|>) to write out its reasoning process, similar to how a human detective might jot down clues before solving a case.
• It Can Use Tools: It can even call external tools (like a scanner) to help investigate, using a special "tool call" token.

The Big Hiccup: The "Blind" Detective
Here is where the story gets real. The researcher trained this model on about 16 million tokens of data (a mix of security images and text). They expected it to learn how to look at a screenshot of a security tool and explain what's happening. Instead, they found a disappointing result: the model mostly ignores the images.

When tested on 50 different security screenshots, the model could only correctly identify the tool being used in about 8% of cases (0.08 score). It could write fluent Spanish sentences, but if you asked it what was in the picture, it often just guessed or hallucinated, ignoring the visual content entirely. The author calls this a "negative result," which is a fancy way of saying, "We tried, and it didn't work yet, and here is exactly why we think that happened."

The Mystery: A Glitch in the Wiring
The paper digs deep into why this happened. They discovered two main things:

1. Not Enough Practice: The model simply hasn't seen enough examples of security images to learn how to connect the dots. The current training volume is too low.
2. A Sneaky Bug: Early on, a software bug caused the model to load with random, untrained weights instead of the smart brain they had built. This made it look like the model had "collapsed" or failed completely, but once they fixed the bug, the model was actually working—just not good enough at seeing yet.

The Big Question: The "NoPE" Puzzle
The most exciting part of the paper isn't the failure, but the mystery it leaves behind. The model's brain uses a special trick called NoPE (No Positional Encoding) on every fourth layer. Most AI models use a system that tells them the order of words or pixels (like "first," "second," "third"). NoPE tries to learn the order implicitly.

The author wonders: Does this NoPE trick help or hurt the model when it looks at images?

• Hypothesis 1: Maybe NoPE is great because images (like a grid of pixels) don't have a strict "left-to-right" order, so forcing an order might confuse the model.
• Hypothesis 2: Maybe NoPE is bad because the model was trained on text (which has a strict order), so it gets confused when looking at a jumbled grid of image pixels.

The paper releases the "recipe" and the "test" for this mystery, inviting other scientists to run the experiment and see which hypothesis is true. They haven't solved it yet, but they've handed everyone the tools to find out.

What's Next?
The author is very clear: This is not a finished product. The model is functional and can run offline, but it isn't ready to replace a human analyst because it can't reliably "see" the screen yet. They plan to run more training with more data and a different learning strategy to fix the vision problem.

In the meantime, they have released everything: the code, the training data, the benchmarks, and the model itself. They are essentially saying, "We built a cool, tiny, Spanish-speaking security detective, but it's currently blind. Here is the blueprint, the test questions, and a specific question about how its brain is wired. We hope you can help us teach it to see."

This approach is refreshing because it prioritizes honesty over hype. Instead of pretending the model is perfect, they admit the failure, explain the technical reasons, and turn the problem into an open challenge for the community. It's a reminder that in science, knowing what doesn't work is just as valuable as knowing what does, especially when you're trying to build something as critical as a tool for keeping digital secrets safe.