Local models

Which models run well on your machine. Measured, not looked up.

What decides how fast a model generates is how fast your memory is, and nearly every tool reads that out of a table of known cards. A table is wrong in a specific way: an RTX 5070 Laptop has a 128-bit bus where the desktop card sharing that number has 192-bit, so keyed on the model number it overstates the laptop by up to 1.8×, and the error lands in every tokens-per-second figure built on it.

WhatRunsHere keeps no such table. It reads the bus width and memory clock off the device, times your system memory in about a second, and marks every figure it shows with where that figure came from.

Every figure above was computed by WhatRunsHere's own solver for a machine of that shape, at 8k context, not typed in by hand. The machines are described from published specifications, so these are estimates, and the application says so on your machine too until you measure it.

weights
attention cache
activations and runtime
headroom left free

Download

Free, and it stays free

One window. It reads the machine it is on, ranks 110 models against it, and fetches the weights when you ask it to. Nothing else leaves the computer: no telemetry, no account, no licence check anywhere in the code.

That is a commitment rather than a stage. A project that never promised to stay free and later charges has broken nothing; one that promised and then charged has broken the only thing it had. So the promise is made here, in public, where you can hold it to account.

No binaries have been published yet. The first release has not been cut, so for now it is built from source: a Rust toolchain and Node 22, about two minutes.

git clone https://github.com/EuBa-Code/WhatRunsHere
cd whatrunshere
npm --prefix ui ci && npm --prefix ui run build
cargo run --release -p whatrunshere-app

Why the answers are different

Three things it refuses to guess

Streaming bandwidth 48.7 GB/s

One thread 24.3 GB/s

Per-token overhead 2.50 ms

It measures your machine, in about a second

Generation speed is set by how fast weights can be read, not by how fast the processor is. WhatRunsHere times that directly rather than looking your hardware up in a table, which is also why it works on a card nobody has catalogued yet, and why it cannot be fooled by a laptop chip that shares its model number with a desktop one.

Weights10.9 GiB

Attention cache104 MiB

Activations51.4 MiB

Runtime400 MiB

Headroom352 MiB

It counts everything the model needs, not just the file

The attention cache is what decides how long a conversation can get, and it is computed from each model's real architecture: grouped-query, sliding windows, DeepSeek's latent attention all have genuinely different slopes. Treating them alike is how a 671B model gets reported as needing more cache than a 70B one when it needs less.

Measured here 48.7 GB/s

From a specification 1008 GB/s

Assumed 30.0 GB/s

It tells you which numbers it actually knows

Every figure carries a rule beneath it: solid for something timed on your machine, dashed for a manufacturer's claim, dotted for an assumption nothing has replaced yet. Sixteen curated models are left out of the catalog entirely because their architecture cannot be described correctly. A list that tells the truth about 110 models is worth more than one that is wrong about 126.

Checked against reality

The models are validated, not argued for

1.5% mean error between the computed weight size and the real file, across 1314 published builds

0.88 to 0.99 R² of the speed model fitted to 181 measurements from 9 real machines

153 tests, none of which touch the network, run on Windows, macOS and Linux

3 constants in the code that came from those measurements rather than from reasoning, each citing the data