26 articles · newest first
12GB is the entry floor for local AI — enough to run genuinely useful models, but with a real ceiling. Here's exactly what 12GB gets you, and when you'll want more.
Ravi Malhotra · Aug 5 · 10 min
The cheap used pick versus the newer 16GB card. More VRAM and speed, or the lowest price? Here's the honest budget-GPU showdown for local AI.
Ravi Malhotra · Aug 5 · 10 min
12GB of VRAM runs more than you'd think — from snappy 7B models to a DeepSeek R1 distill. Here's exactly which local models the RTX 3060 12GB can run, and how fast.
Ravi Malhotra · Aug 5 · 10 min
Years on, the RTX 3060 12GB is still the cheapest sensible entry to local AI — 12GB of VRAM for around $200-250 used. Here's honestly how good it still is, and the one price catch.
Ravi Malhotra · Aug 5 · 10 min
The o1-mini-beating 32B distill wants ~20GB of VRAM, so 24GB is the target. Here are the best GPUs to run DeepSeek R1, from the value king to the fastest.
Ravi Malhotra · Aug 5 · 10 min
Two cheap 16GB cards give you 32GB of VRAM — enough for 30B+ models — for around the price of a single high-end GPU. Here's whether that dual-5060-Ti path actually makes sense.
Ravi Malhotra · Aug 5 · 10 min
They run the same models — but one costs nearly twice as much. For local AI, where VRAM matters more than raw speed, the cheaper card makes a surprisingly strong case.
Ravi Malhotra · Aug 5 · 10 min
16GB of VRAM opens up a lot — from snappy 7B models to gpt-oss-20b. Here's exactly which local models the RTX 5060 Ti 16GB can run, and how fast, with real benchmark numbers.
Ravi Malhotra · Aug 5 · 10 min
At $549 it's the cheapest new way to get 16GB of VRAM — and for local AI, VRAM is what matters. Here's honestly how good the RTX 5060 Ti 16GB is, and where its limits are.
Ravi Malhotra · Aug 5 · 10 min
Qwen3-Coder's best variants want 24GB of VRAM, and its huge context wants headroom. Here are the best GPUs to run it, from the value king to the fastest option.
Ravi Malhotra · Aug 5 · 10 min
gpt-oss-20b fits on a 16GB card, but the model's huge context can quietly overflow it. Here are the best GPUs to run it, from the cheapest 16GB option to the 24GB value king.
Ravi Malhotra · Aug 4 · 10 min
Building a local-AI rig and unsure how big a PSU to buy? Here's the simple wattage math for single and dual GPUs — plus why power-limiting lets you get away with a smaller supply.
Ravi Malhotra · Aug 4 · 10 min
Because AI inference is memory-bandwidth-bound, the last chunk of your GPU's power budget buys almost no extra tokens. Cap the wattage and you run cooler, quieter, and cheaper for basically free.
Ravi Malhotra · Aug 4 · 10 min
eGPUs get a bad rap from gaming benchmarks — but local AI is a totally different workload, and an external GPU can hit ~92% of native speed. Here's whether one's worth it for your laptop.
Ravi Malhotra · Aug 4 · 10 min
GDDR7 modules and TSMC wafer costs just pushed every RTX 50 card higher this week — but the ugliest number isn't the one everyone's quoting.
Ravi Malhotra · Aug 4 · 6 min
AMD's Strix Halo mini PCs put up to 128GB of unified memory in a box the size of a book — and that lets them run models no single consumer graphics card can hold. Here's the pick, and the catch.
Ravi Malhotra · Aug 4 · 10 min
Nvidia still owns local AI because of CUDA. But AMD's ROCm has quietly grown up — and in 2026 it's a genuine option, with caveats. Here's the honest state of Radeon vs GeForce for running LLMs.
Ravi Malhotra · Aug 4 · 10 min
Everyone asks about VRAM — but your system RAM matters too, especially without a big GPU. Here's how much RAM you actually need for local AI, from 16GB to 128GB, and what each tier runs.
Ravi Malhotra · Aug 4 · 10 min
The Tesla P40 is the cheapest way to get 24GB of VRAM, period. But it's a headless server card with no fan, and it's slow. Here's whether the bargain is worth the hassle.
Ravi Malhotra · Aug 3 · 10 min
Both have 16GB of VRAM, so for local AI they run the same models. The 5080 is faster, but for AI that matters far less than you'd think. Here's the honest call — and a better idea.
Ravi Malhotra · Aug 3 · 10 min
Copilot+ 'AI PCs' have a dedicated NPU chip — but here's the catch nobody advertises: it can't run local chatbots. A cheap old GPU runs them 10x faster. Here's the truth.
Ravi Malhotra · Aug 3 · 10 min
Windows is the easy path and works great for most people. Linux gives power users an edge, especially with AMD. Here's which OS to run local AI on, and why it matters less than you think.
Ravi Malhotra · Aug 3 · 10 min
If your local model feels sluggish, it's usually one of a few fixable things. Here's how to speed up local LLM inference, from the one change that matters most to the fine-tuning.
Ravi Malhotra · Aug 3 · 10 min
Apple Silicon's unified memory makes Macs quietly great for local AI: the RAM number is how big a model you can run. Here's which Mac to buy for AI, from the Mini to the Studio.
Ravi Malhotra · Aug 3 · 10 min
A second GPU doubles your VRAM, which unlocks 70B models — but it's not free performance, and most people don't need it. Here's when a dual-GPU local-AI rig is actually worth it.
Ravi Malhotra · Aug 1 · 10 min
A local coding model gives you a private, offline, unlimited Copilot — but the good coding models are big. Here's the GPU you need to run them well, at every budget.
Ravi Malhotra · Aug 1 · 10 min