benchmark local llms.no |

curated stem questions from mmlu — cs, math, science, engineering. runs entirely in your browser via webgpu. compete on a global leaderboard.

9 models from 0.6b to 8b parameters — qwen, llama, gemma, phi, deepseek r1, mistral, smollm2. all running locally via webgpu. no api keys. no server.

qwen3 0.6b0.6B

qwen~400mb

llama 3.2 1b1B

llama~720mb

smollm2 1.7b1.7B

smollm2~1.1gb

gemma 2 2b2B

gemma~1.3gb

llama 3.2 3b3B

llama~1.8gb

phi 3.5 mini3.8B

phi~2.2gb

qwen3 4b4B

qwen~2.5gb

mistral 7b7B

mistral~4.3gb

deepseek r1 8b8B

deepseek~5.0gb

downloads once, cached locally · runs entirely on your gpu

multiple-choice questions from mmlu — cs, math, science, engineering. hard enough that even 8b models score under 60%. no ceiling.

question 1 / 200 correct

csmediumcomputer-science-14

Which of the following sorting algorithms has the best average-case time complexity?

A.Bubble sortB.Insertion sortC.Merge sortD.Selection sort

model response

▋

see exactly where your model fails — broken down by subject and difficulty. click any question to review the full reasoning.

438/ 1000

easy

7/8+0pts

medium

6/8+0pts

hard

4/4+0pts

questions

benchmark local llms.no api key.|