Open-weight models · calculator
What can I run locally?
Which open-weight model to run on your own machine, for the job you pick, and what it costs in power. We guess your hardware from what your browser says and your electricity price from your time zone, and you can change both. 120 graded models, each with the builds that fit.
How it works
- A model fits when the build's memory need, the file plus a fifth for the runtime and a working context, is no more than your memory times the share a model can have (92% of a graphics card, two thirds to three quarters of unified memory, 60% of a computer's RAM).
- The pick is the best-graded model for the job with a build that fits and runs on your kind of machine, in its 8-bit build if that fits, then 6, 5 and 4-bit. A 16-bit build is picked only when nothing quantised fits, since it runs at about half the speed for little gain.
- Speed is an estimate for one stream, from your memory bandwidth and the gigabytes read for each token (the build's size times the share of parameters active at once). Real speed depends on the runtime, the context length and the settings.
- Cost is power under load times hours a day, at your electricity price, and per million tokens at the estimated speed. It leaves out the machine itself.
Nothing your browser reports is sent to us. A browser can tell us a graphics card's name and how many threads the processor has, not how much memory the machine has, so memory is a guess for you to correct.
The machines it knows
Memory and bandwidth are the makers' published figures. Power is a graphics card's board power, or our estimate for a whole machine under load (marked est.). Change any of it above to match yours.
| Machine | Memory | Bandwidth | Power |
|---|---|---|---|
| NVIDIA RTX 3060 12 GB | 12 GB | 360 GB/s | 170 W |
| NVIDIA RTX 4060 Ti 16 GB | 16 GB | 288 GB/s | 165 W |
| NVIDIA RTX 3090 24 GB | 24 GB | 936 GB/s | 350 W |
| NVIDIA RTX 4090 24 GB | 24 GB | 1008 GB/s | 450 W |
| NVIDIA RTX 5070 Ti 16 GB | 16 GB | 896 GB/s | 300 W |
| NVIDIA RTX 5080 16 GB | 16 GB | 960 GB/s | 360 W |
| NVIDIA RTX 5090 32 GB | 32 GB | 1792 GB/s | 575 W |
| NVIDIA RTX PRO 6000 Blackwell 96 GB | 96 GB | 1792 GB/s | 600 W |
| Apple M1, M2 or M3 | 8, 16, 24 GB | 100 GB/s | 30 W est. |
| Apple M1 Pro, M2 Pro or M3 Pro | 16, 18, 32, 36 GB | 150 GB/s | 50 W est. |
| Apple M1 Max, M2 Max or M3 Max | 32, 36, 48, 64, 96, 128 GB | 400 GB/s | 80 W est. |
| Apple M1 Ultra or M2 Ultra | 64, 128, 192 GB | 800 GB/s | 160 W est. |
| Apple M4 | 16, 24, 32 GB | 120 GB/s | 30 W est. |
| Apple M4 Pro | 24, 48, 64 GB | 273 GB/s | 60 W est. |
| Apple M4 Max | 36, 48, 64, 128 GB | 546 GB/s | 90 W est. |
| Apple M3 Ultra | 96, 256, 512 GB | 819 GB/s | 180 W est. |
| Apple M5 | 16, 24, 32 GB | 153 GB/s | 30 W est. |
| Apple M5 Pro | 24, 48, 64 GB | 307 GB/s | 65 W est. |
| Apple M5 Max | 36, 48, 64, 128 GB | 614 GB/s | 100 W est. |
| AMD Ryzen AI Max+ 395 (Strix Halo) | 64, 96, 128 GB | 256 GB/s | 140 W est. |
| NVIDIA DGX Spark | 128 GB | 273 GB/s | 200 W est. |
| Any computer, on the CPU | 8, 16, 32, 64, 128 GB | 80 GB/s | 100 W est. |
Where the defaults come from
Electricity prices are household prices in US dollars a kWh, the 2023 to 2026 averages from GlobalPetrolPrices, read on 9 November 2026, and the world average is its third quarter of 2026 ($0.179). Your country is guessed from your time zone. The models, their grades and their builds are our listings, graded with the open-weight reading of the checklist, and every build's memory need is on its listing. All the models by lab are on Your cousins.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.