The benchmark
Your computer runs an AI model on up to 50 tasks: summaries, facts from an email, exact instructions, reasoning, code, tools and small web pages.
The 16 harder tasks follow for larger models: games and apps to build, code against hidden tests, a query, a puzzle and a long tool chain.
A run takes one of three sizes. Fast is 24 tasks of text answers. Medium adds tools, code and planning, 37 tasks. Full is all 50, with the web pages, games and apps. Only a full run takes a place on the leaderboard.
The site checks every answer itself and times each one. You see what the model said, what it got right and what it built.
A model is the AI itself: a file you download once and run with a free app such as Ollama, LM Studio or llama.cpp. The number in its name, like 4B or 8B, is its size in billions of parameters: bigger is usually smarter and needs more memory.
The score
Higher is better: speed in tokens a second, multiplied by the share of points passed, times 100. Each of the 34 everyday tasks is worth 1 point and each of the 16 harder ones is worth 3, so 82 points in all. A fast or medium run is scored on its own tasks: 24 points for fast, 53 for medium.
A token is about three quarters of a word, so 60 tok/s is roughly 45 words a second. Tiers: Bronze below 1,500, Silver from 1,500, Gold from 3,000, Platinum from 6,000, Diamond from 10,000. A tie goes to more points passed, then more tasks passed, then the earlier run.
A score ranks on the leaderboard when a signed-in account posts a full run with its computer named (from the picker, or typed in), answers every task through a model's API, and its speeds look measured. Each account keeps its best score per computer. Up to ten runs a UTC day per account can rank, across its computers and models. Only runs through a model's API count toward those ten: every one posted with a computer keeps its place, deleted or not, so a run cannot be swapped out for another. No more than five runs of one model on one computer a UTC day can rank. The rest are graded and shown, with the reason. A run that does not rank (a task left out, no measured speed, a speed the computer cannot reach, the UTC day's limit, or an agent's direct answer) comes after the ranked runs. An agent's direct answers are never ranked on the leaderboard. All runs lists every run, ranked or not.
How speeds are checked
The browser or agent that ran the model measures and reports speeds, so they are checked, not proven. A speed the computer cannot reach for the model's file is held for review, as is one far above the same machines and one repeated on every task. The file's size comes from the listed file. Without one, it is estimated from the largest size in the model's name (a 7B model is about 4 GB at 4.5 bits a weight, and 8x7B counts as 56B), and never less than 100 MB: a name with no size, or a small one, is checked as a 100 MB file, and so is a listed file that small. A computer is checked at the memory bandwidth of the listed computer whose words its name contains, wherever in the name they sit, so "RTX 3060 OC" is an RTX 3060 but "RTX A1000" is not an A100. A name with a CPU's words (Ryzen, Xeon, Core i7, Celeron, an Intel N-series or an AMD G-series chip) is checked as a CPU. A name that contains no listed computer is checked against the fastest listed computer of its own kind: a GPU against the fastest GPU in the list, an Apple chip against the fastest Apple chip, and a CPU at 600 GB/s. Anyone can flag a run.
Which computers count as the same
The same chip or graphics card, whatever its memory. Names ignore capitals, spaces, hyphens and brand words, so "GeForce RTX 3060", "RTX-3060" and "rtx 3060" are one computer. Memory does not split a computer: a card typed with 8 GB and with 12 GB is one computer on the board, which shows the memory of the account's best run on it. A memory size typed into the name ("RTX 3060 12GB OC") does not split it either. A name typed another way gets no more of the UTC day's ten API runs, which count across all of an account's computers and models.
Runs and comments
Runs, comments and challenge proposals appear at once. Anyone can flag one: a flagged run is hidden until an admin checks it, and a flagged comment stays up, marked, until then. You can delete your own runs and comments. Older run reports from before the benchmark stay as history.
Members can vote up the pages models built, and votes never change a score.
Takedowns
Send takedown requests to admin@y3k.xyz. Within 72 hours we remove what was named.
Privacy
No account is needed. An account is a username and password, with no email, so there is no password recovery. Your chosen computer is remembered until you close the browser, or 30 days with Remember on this device. Keep private details out of runs, notes and comments.
What is stored
An account stores a username and a scrypt password hash. It also keeps a hash of your API token if you make one, a failed sign-in count and lock time, when you last visited, and up to 20 saved hardware setups.
If you sign in to the chat, the chat server also keeps an account with the same name. Opening the chat page while signed in stores a hashed 12-hour chat passphrase against your username, so the chat can sign you in. Changing your password ends it.
The chat page shows who is in the room and the last five messages to anyone who opens it. Anonymous runs, flags and proposals are rate-limited by a salted hash of your address, kept for 24 hours.
Logs
Hosting logs may include request metadata.
