What are 1B, 3B, 7B and 14B?
B means one billion parameters: 7B means roughly 7 billion numerical values learned during training. More parameters generally need more memory, but do not guarantee better answers.
BETA
Braintec AI Calculator
Would you like to discover which AI can run on your computer?
No account and no software to install. Your computer information is processed in the browser and is not sent to the server.
PRACTICAL GUIDE
Understanding local AI, without being an expert
This tool does not download or run AI: it estimates which model sizes may be considered based on your computer memory. Here is what the terms and steps mean.
What is Q4, Q5 or Q8 quantization?
It is a way to store model weights using fewer bits. Roughly, Q4 uses about 4 bits per weight, Q8 about 8, and FP16 uses 16. Fewer bits reduce weight memory but may affect accuracy. Real files also contain additional information.
Q4 less memoryQ8FP16 more memory
Why do we ask about memory?
A model needs memory for its weights and, while running, for the conversation (KV cache), buffers and the application running it. If available memory is unknown, 70% of installed RAM is only a starting estimate, not a measurement or guarantee.
How to read your result
Light, Balanced and Advanced are three size classes to consider, not three programs to install. The current calculation checks only the theoretical minimum for weights: actual execution also depends on file format, extra memory and runtime compatibility.
What do the detection commands do?
The commands in step one read information already available on the system. They install no software, require no administrator rights and do not change settings. You can inspect the command before running it and review its output before pasting.
Windows · PowerShell
Get-CimInstance reads system information via CIM/WMI: total memory, system-reported free memory, CPU and GPU name. Write-Output prints the result.
macOS · Terminale
sysctl reads hardware properties such as installed RAM and architecture. printf formats the output; sed adds labels. This command does not measure available memory.
Linux · Terminale
The /proc/meminfo and /proc/cpuinfo files expose kernel information about memory and CPU. awk extracts only relevant fields; printf marks the output.
Important
GB and GiB differ: 1 GB = one billion bytes; 1 GiB = 1,073,741,824 bytes. Weight memory is only part of total requirements. The calculator does not currently verify speed, GPU support or individual model compatibility.
Technical details for professionals
Weights and numeric precision
Theoretical weight-only estimate: parameters × effective bits per weight ÷ 8. Convert bytes to GiB by dividing by 1,073,741,824. Q4 and Q8 labels alone do not define effective bits per weight: metadata, scales and blocks introduce differences.
Context and KV cache
Context token count affects inference memory. Estimating KV cache requires architecture data such as layers, KV heads, head dimension and cache precision. KV heads are groups of Key and Value vectors per layer; head dimension is the number of elements in each vector. KV cache is temporary memory used while processing context, not permanent conversation memory. Without these details, a reliable value cannot be determined.
Unified and dedicated memory
Unified memory is shared by CPU and GPU and must not be counted twice. With dedicated VRAM, GPU capacity, system RAM and runtime offload support must be assessed separately: they are not one freely additive pool.
Interpretation and limitations
The 70% RAM assumption is not a measurement. Passing the weight check does not prove compatibility, speed, CPU/GPU instruction support or model quality. No tokens/s performance is claimed without measurements.
Discover your computer
Select your operating system. We will show a read-only command to collect only relevant hardware characteristics.
The command reads general processor, memory and graphics information. It does not collect usernames, serial numbers or network addresses. Always review what you paste.
I prefer the system settings
In graphical mode, you may enter available memory below. Automatic screenshot recognition is not available in this beta.
Not comfortable with a terminal? Skip this step and enter RAM manually. The command simply helps find the values.
RAM: memory used by the operating system and applications. Available memory: the portion the system can allocate to new tasks. VRAM: dedicated graphics memory; do not automatically add it to RAM.
If available memory is unknown, 70% of installed RAM will be assumed. This is an estimate, not a measurement, and does not guarantee a model will run.
How would you like to use AI?
This choice helps us explain possibilities without attributing unverified abilities to models.
Your choice describes what you want to do, not your hardware. A larger model is not guaranteed to understand images or write code.
Model configuration (optional, advanced)
These values describe a hypothetical model, not a specific product. Effective bits per weight must come from format documentation or the model file.
Your assessment
Memory details
How we calculated this result
Beta version: assesses memory, not runtime compatibility or performance. No tokens/sec are estimated. GGUF formats and specific quantizations require verified metadata.