Will it run? Know before you buy.
Pick your hardware or budget and see which local LLMs actually run, at what speed, and what to buy instead.
Free while in beta. No spam — just the launch link and benchmark updates.
Why Local LLM Hardware Decisions Matter Right Now
Local language models have quietly crossed a threshold. A 70B model running at Q4 quantization on a machine you own can now handle real work — drafting, coding, analysis — without sending a single token to a cloud provider. The catch is that the hardware to run it well costs anywhere from two to eight thousand dollars, and the difference between a machine that hits 40 tokens per second and one that crawls at 3 tokens per second is often invisible on a spec sheet. Memory bandwidth, VRAM capacity, and quantization level interact in ways that no product page explains.
The current conversation proves the anxiety is real. When Llmspeed launched on Hacker News, it drew hundreds of points and dozens of comments — but almost every comment was a variation of the same question: "My 4090 runs this model at this speed, is that normal?" People are trading benchmark numbers in forum threads because there is no structured way to answer the question before spending money. Reddit threads, Ollama's model pages, and vendor requirement lists all tell you roughly what fits, but none of them tell you what performance to expect, or whether a cheaper configuration would serve you better.
That gap is exactly why a pre-purchase calculator matters today. The community already has the data — scattered across comment sections and forum posts — but it is not queryable, not comparable, and not tied to a budget. WillItRun turns that scattered experience into a single page: enter your hardware or your budget, get a clear green-yellow-red verdict, an expected tokens-per-second estimate with the formula shown, and a short list of alternatives worth considering. It is the answer people are already hunting for in comment threads, delivered before the credit card comes out.
How It Works
Describe your setup or your budget
Choose from common rigs like an RTX 4090, a Mac Studio with unified memory, or an AMD Strix Halo, or enter a custom VRAM and memory bandwidth figure. If you have not bought yet, just enter a budget and we work backwards from there.
Pick the models you care about
Select from popular local models such as Llama 3.3 70B, Qwen2.5 32B, Mistral Large, or DeepSeek-R1 distills, and set a quantization level. We estimate fit and speed using memory bandwidth divided by model size, with the efficiency factor shown openly so you can challenge the math.
Read your report and share it
Get a green, yellow, or red verdict with the bottleneck named, an expected tokens-per-second range, a side-by-side comparison of alternatives at the same budget, and a shareable link you can drop into a forum to ask for a second opinion.
What You Get
Clear run verdicts
Every model gets a green, yellow, or red rating with the actual reason attached — insufficient VRAM, a memory bandwidth ceiling, or a quantization mismatch — so you know whether the limit is capacity or speed.
Expected tokens per second
We estimate throughput from memory bandwidth and model size, then show the formula and the efficiency coefficient in plain sight. You can see exactly where the number comes from and decide how much to trust it.
Shareable comparison links
Each report generates a permanent URL encoding your hardware and model choices, so you can paste it into a forum thread and let people with real rigs confirm or correct the estimate before you commit.