Your "old" PC isn't dead. Find out what it can still do.
Enter your RAM, processor and graphics card for an indicative estimate of compatibility, local models and Linux distributions. Calculations run in your browser and the values you enter are never sent anywhere.
1️⃣ Your PC
Don't know your exact processor? The family works too: 4th gen Core i5, Ryzen 5 3600, Pentium 4…
Type a few letters of the name, then pick from the list
Total physical memory, not the "usable" amount
The estimate uses a simplified budget for the system: real usage changes with version, drivers and programs. It's selected automatically for Apple processors
Not in the list — I'll enter it manually
The GPU can speed up inference if the runtime and backend support it; VRAM doesn't guarantee the whole model will fit.
An SSD can reduce loading times compared to an HDD; results vary by drive and system
Syncs to the CPU's launch year; the PC may have been built or bought later.
🎯 The verdict on your PC
—
0/100
Indicative index, not a benchmark: it combines RAM (45%) and processor (55%). The year stays as context and doesn't count toward the score; compatibility, models and speed must be checked separately.
Is Linux worth it?
—
LM Studio
—
Windows 11
—
—Level
—Usable free RAM
—Max parameters
—Models within estimate
—7B Q4 speed (reference)
2️⃣ Models to evaluate on your PC
—
Model
Parameters
Quant.
File
Estimated RAM
Estimated speed (Q4 base)
Verdict
Select year, RAM and processor, then press Calculate.
How to read the table. The verdict applies indicative internal criteria and doesn't guarantee the model will start. Estimated RAM considers weights and a reference context cache: the OS, apps and configuration can require more memory. Estimated speed (Q4 base) is an indicative range in tokens per second, not a benchmark; other quantizations can change the result. Quant. indicates a weight format: size and quality vary by model and task.
3️⃣ Compare two PCs
Compare two setups for Linux, LM Studio, speed and models. The result is an estimate and a simple category count: it doesn't weight the tasks that matter most to you and doesn't replace a real test.
A
VS
B
Item
PC A
PC B
Winner
4️⃣ Which Linux to install
—
Watch out for 32-bit. Installable options are few and support varies: antiX 26.1 and Debian 12 i386 are possible candidates, but check CPU instructions, releases and drivers on the official sites. Debian 13 doesn't offer an i386 kernel or installer; Debian 12 i386 requires at least an i686 CPU and receives LTS updates until June 2028.
Windows: press Ctrl+Shift+Esc and open Performance → CPU and Memory, or press Win+R, type msinfo32 and read Processor and Installed RAM. The year field indicates the computer's era, not its speed. Linux: open a terminal and run lscpu (processor, cores, SIMD support) and free -h (RAM). No software needed: the label on the motherboard inside the case, or the code printed on the CPU, shows the exact model.
Does LM Studio work on any processor?
For Windows x64, LM Studio's official requirements indicate AVX2; on Linux x64 the package includes AVX2 by default. Check the requirements of the platform and build you choose: an alternative runtime may have different requirements, but the calculator doesn't guarantee its compatibility. Apple Silicon uses a native ARM path.
Why does the model need more GB of RAM than its parameters?
Besides the weights, memory is needed for the KV cache, runtime buffers, the context and the operating system. The cache grows with context and depends on the model architecture. As a rule of thumb, leave headroom over the file size and check free memory before loading it.
What's the difference between Q4, Q5, Q8 and "MXFP4"?
They are quantization formats that take up less space than higher-precision weights. Q4_K_M is a common memory/quality trade-off, but the effect varies by model and task; it isn't indistinguishable from the original in every situation. Q8 needs more memory and can preserve more precision, while heavier quantizations can penalize some tasks more.
Is a 2012 PC with 16 GB better than a 2019 one with 8 GB?
It depends on the model and the use. More RAM lets you load bigger models or use larger contexts, but a bigger model isn't automatically better at every task. CPU, GPU, memory bandwidth, quantization and backend affect speed. Dual channel increases the available bandwidth, but the gain varies with the platform and workload.
Is Linux worth installing on a PC that still runs Windows?
It can be worth it, but check the distribution's support and compatibility with your machine. Windows 10 ended mainstream support on October 14, 2025. In the European Economic Area, eligible users can join the free ESU program by staying signed into the same Microsoft account; the option for local-account users is a one-time purchase. The program ends on October 12, 2027. Windows 11 compatibility depends on the exact CPU model and requirements such as TPM 2.0, UEFI and Secure Boot: the generation alone isn't enough. Try Linux in Live mode and check drivers, peripherals and applications before installing.
Can I use the PC as an AI server? Like running the model on an old laptop and asking it from another PC?
Yes. Ollama can provide a local API; to connect another PC on the network you must configure network listening and the firewall, because the API is local to the computer by default. Limit access to your trusted network and don't expose it directly to the Internet. Inference speed depends mainly on the computer running the model; the client handles the interface and the connection. Read the guide to connecting multiple PCs before changing your network.
How to evaluate and revive an old PC
A 2008-2012 PC can still be suitable for some tasks, but there's no verdict that works for everyone: component condition, operating system, CPU, RAM and required software all make a difference. Small local AI models can run on some setups; performance and compatibility must be verified on the device.
The three things that really matter
1. RAM limits how large the models you can load are, but weights, context cache, system and open programs consume memory. Always check the model's requirements and leave headroom.
2. CPU, GPU and runtime affect performance and compatibility. SIMD requirements vary by platform: LM Studio indicates AVX2 for Windows x64 and includes it in the default Linux x64 package; other runtimes may have different configurations.
3. Memory speed can affect CPU inference. Dual channel increases theoretical bandwidth, but the real benefit depends on CPU, memory, model and backend.
What "running AI locally" means
With a local runtime, the model runs on your own computer; requests and files don't have to be sent to a remote service, except for online features or telemetry in the app you choose. Compatibility, privacy, speed and quality depend on the runtime, the model and the configuration. Some models also support images or other inputs, but they require a compatible runtime.
Why the processor and not the GPU
A compatible GPU can speed up inference if the runtime supports it and enough of the model fits in VRAM. The effect varies a lot by GPU, backend, quantization and context; a fixed speedup number isn't reliable for every PC. Integrated graphics share system RAM, and acceleration support depends on model, driver and runtime. Before buying a card, check VRAM, power supply, space, firmware and software support.
The role of Linux
Linux can be useful to extend the life of a PC or try a different environment, but support depends on the distribution and the compatibility of drivers, browsers and peripherals. Memory usage changes with version, desktop, updates and active services. AI app compatibility is a separate check: look at the requirements of the runtime you choose, especially the SIMD instructions required by the CPU.
Settings to try
Threads, context window, quantization and GPU offload affect memory and performance, but no combination improves every PC by the same amount. Start from the default settings, reduce the context or model size if memory is short, and compare the results with a real test. Swap and zram can help stability, but they don't replace enough RAM and can slow down inference.
The right time to recycle your PC
To decide whether to revive it, first check the disk condition, the maximum RAM supported by the motherboard, operating system compatibility and the applications you need. Age helps you get oriented, but it doesn't determine performance or compatibility on its own. An old PC can be useful for light tasks or as a remote client, while running models locally depends on CPU, memory, GPU and runtime.