I wanted one honest answer.
Which local AI models can my computer actually run well?
VRAM Check began because every answer I found stopped just before the part that mattered: what the experience would really feel like on the hardware I already owned.
The original frustration
I was trying to use local AI on a modest ASUS ROG STRIX GeForce GTX 1080. Tools such as Ollama and LM Studio made running models approachable, and their catalogs gave me a general idea of what might fit. But “might fit” was not the answer I needed.
Some suggested models would not run properly on my setup. Others technically worked, but a basic question felt like it took forever. I spent more time comparing model sizes, quantizations, memory requirements, and scattered opinions than actually using local AI.
If I had this question, other people probably had it too.
That simple thought became the reason to start. Not because the idea was grand, but because the problem was persistent, practical, and surprisingly hard to answer with confidence.
From a recurring question to a measured answer.
The product did not arrive fully formed. Each version removed one assumption and replaced it with something observable.
The question
What can my computer actually run?
I could find model lists and broad recommendations, but not a dependable answer for my own machine. A model could technically load and still take so long to answer a basic question that it was not useful in practice.
The pattern
The same doubt kept appearing.
While researching, I kept finding versions of the same question in Reddit discussions and local AI communities: Which LLM can my GPU run, and will it feel fast enough to use? It was clearly not only my problem.
The first attempt
VRAM Check began as a manual tool.
The earliest versions asked people to enter their hardware and returned a rough estimate. That was useful for exploring the idea, but it still depended too heavily on specifications and assumptions.
The turning point
The computer needed to answer for itself.
The project changed when the question became: can one shared benchmark measure the real machine, preserve the evidence, and turn the result into guidance that a person can actually understand? That question became the foundation of the CLI and the public results you see today.
Built honestly, one hard lesson at a time.
I knew how to shape the product. Learning how to build it was part of the work.
My name is Kenneth Michel Chaves López. I am a SaaS technology director with experience in process architecture, UX/UI, and leading software products. I understood how to define the problem, design the experience, and decide what a useful product should do. Turning that vision into a working benchmark, service, CLI, and public catalog was a different challenge.
I used artificial intelligence tools to explore technical possibilities and help close that implementation gap. The first versions were basic and completely manual. Repeated testing, failed assumptions, strange hardware results, backend problems, and feedback from real runs gradually shaped a more serious system.
VRAM Check is still in progress. Every new generation of models and GPUs creates another reason to improve the catalog, question the scoring, and make the recommendations more useful. I do not think that work ever becomes completely finished, and that is part of the point.
Connect with Kenneth on LinkedInWhat the project is trying to protect.
The implementation will keep changing. These three ideas should not.
Measure before recommending
A specification can describe a GPU. A measured run shows what the complete computer did under the same workload used for every comparable result.
Useful matters more than possible
Loading a model is not enough. The experience also needs enough memory, reasonable response speed, and a stable run to be practical.
Show where certainty ends
Measured performance, estimated model fit, and upgrade guidance are different claims. VRAM Check is being built to keep those boundaries visible.
The question is still simple.
Run the benchmark on your own computer and leave with an answer grounded in what it actually did.