measuredlocally.comLeaderboard Round one, 15 to 20 October 2026
Local LLM model leaderboard
Eight tests of the AI models you can run at home, on the machines I run them on. Round one has four models.
The leaderboard
| Model | Ran on | Games | 3D | SVG | Frontend | Animation | Dashboard | Debugging | Reasoning | Overall | All eight took |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Nemotron 3.5 Bolt 74B | RTX 4090 desktop | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 83 | 97.9 | 3.2 min |
| GLM 5.3 Flash | Four DGX Sparks | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 83 | 97.9 | 7.5 min |
| DeepSeek V4.1 Flash 1017 | Four DGX Sparks | 83 | 100 | 100 | 100 | 100 | 100 | 100 | 83 | 95.8 | 4.1 min |
| Qwen 3.8 27B | RTX 4090 desktop | 33 | 50 | 100 | 100 | 100 | 62 | 83 | 67 | 74.4 | 16.7 min |
Each test's score is the share of its checks passed, and the overall is the average of the eight. The time is for all eight tests, one after another, on the machine named. My Sparks would run a little faster than these figures. I cap their graphics chips at 2,200MHz, against about 2,400 uncapped, to cut their power use. When I tested the cap on one box in August, writing speed stayed within 1 per cent, reading was 2 to 4 per cent slower, and the chip's peak power fell from 55.5W to 43.3W.
The eight tests
| Test | What it asks for |
|---|---|
| Breakout Games | A Breakout game in one web page: five rows of bricks, a paddle and a ball, moved with the arrow keys, the mouse or a finger, with a score, three lives, and screens for game over and for winning. |
| Lighthouse at dusk 3D | A 3D scene made with three.js: a lighthouse on a rocky island at dusk, its beam sweeping round, waves on the sea, stars coming out, and a camera that circles on its own and turns when you drag. |
| Fox in the snow SVG | A picture drawn in code (SVG): a red fox sitting in deep snow, birch trees behind it and snow falling, made of at least 25 shapes and no images. |
| Landing page Frontend | A landing page for a made-up product, a quiet box that runs AI at home: a headline and a button, three feature cards, three prices in pounds, and four questions that open when clicked. |
| Balls in a spinning heptagon Animation | Twenty numbered balls falling and bouncing inside a seven-sided shape that spins, with gravity, friction and bounces off each other. No ball may ever escape. |
| Sales dashboard Dashboard | A sales dashboard from a small table of six months' sales of three products: a line chart, a bar chart, three summary figures worked out from the data, and a box to tick for each product. |
| Three bugs in a small module Debugging | A short piece of JavaScript with three bugs in it, to find and fix. |
| Six questions with exact answers Reasoning | Six short questions with one right answer each: times, sums, dates and a counting puzzle. |
How I test
- The same prompts for every model, kept private. Public prompts end up in what new models are trained on, so the exact words are not here; their fingerprint (version r1, 2f0794a2dc4cd21e) shows a later round used the same ones.
- One attempt at each test, thinking off, the sampling settings the engine or the model's maker gives, and room for 16,000 tokens.
- Every build runs in a real browser, with the network off. A test's score is the share of its automatic checks passed, from 0 to 100.
- Quality does not depend on the machine; speed does. How fast each machine runs each model is on the hardware leaderboard.