Skip to content
AI Speedtest
← Back to benchmark

How the benchmark works

The test runs fixed language and embedding workloads on your browser’s WebGPU adapter and reports the median of three or five measured passes.

Measurements

Text generation tok/s
SmolLM2 generates 128 output tokens using greedy decoding with seed 42 and end-of-sequence stopping disabled. The first token comes from prompt processing, so throughput covers the remaining 127 decoding steps.
Prompt processing tok/s
The model reads a fixed prompt using a fresh KV cache for every pass. Token counts come from the model’s tokenizer.
Time to first token ms
WebLLM reports the time spent processing the prompt before producing the first token. This measures runtime prefill latency rather than latency through a streaming interface, and lower values are better.
Embeddings tok/s
Arctic Embed processes four fixed documents into 384-dimensional vectors. Throughput includes tokenization and output transfer, and the output is checked for finite, nonzero values.

AI score

The score combines the median throughput of each workload, with text generation weighted at 50%. Higher scores indicate faster performance on this test.

1000 × (decode / 50)^0.50
     × (prefill / 500)^0.25
     × (embedding / 1000)^0.25

A score of 1,000 corresponds to chosen targets of 50 generation tok/s, 500 prompt tok/s, and 1,000 embedding tok/s. These targets are not measurements of a reference computer, and the score is neither a percentile nor a hardware ranking.

Test conditions

Each model completes one unscored warm-up before three or five measured passes. Downloads, model loading, and shader compilation during warm-up are excluded from the results, and models load sequentially to reduce peak GPU memory.

Keep the tab visible and close competing GPU workloads for a more consistent result. Compare runs using the same protocol and model versions, since the browser, graphics driver, power mode, and device temperature can affect performance.

The test measures inference through the browser’s WebGPU adapter; it does not measure dedicated NPUs, native CUDA performance, training speed, model intelligence, or the largest model your device can run.

Models and inputs

Language model
SmolLM2-360M-Instruct-q4f32_1-MLC

The model uses 4-bit weights, float32 computation, and a 2,048-token context.

Embedding model
snowflake-arctic-embed-s-q0f32-MLC-b4

The model uses float32 weights and computation with a batch size of four.

Model and runtime files download from Hugging Face and GitHub on first use and are cached when browser storage permits. Reported hardware details are limited to what the browser exposes; storage buffer limits do not represent GPU memory capacity.

Benchmark inputs
Summarize the following notes and explain their practical implications in detail.
Local artificial intelligence processes data on a personal computer. Language models generate text one token at a time. Prompt processing handles the input before generation begins. Embedding models turn documents into vectors for semantic search. Performance depends on the model, precision, memory, browser, and graphics driver. A repeatable benchmark uses fixed inputs and excludes model downloads and warm-up.
Local artificial intelligence processes data on a personal computer. Language models generate text one token at a time. Prompt processing handles the input before generation begins. Embedding models turn documents into vectors for semantic search. Performance depends on the model, precision, memory, browser, and graphics driver. A repeatable benchmark uses fixed inputs and excludes model downloads and warm-up.
Local artificial intelligence processes data on a personal computer. Language models generate text one token at a time. Prompt processing handles the input before generation begins. Embedding models turn documents into vectors for semantic search. Performance depends on the model, precision, memory, browser, and graphics driver. A repeatable benchmark uses fixed inputs and excludes model downloads and warm-up.
Local artificial intelligence processes data on a personal computer. Language models generate text one token at a time. Prompt processing handles the input before generation begins. Embedding models turn documents into vectors for semantic search. Performance depends on the model, precision, memory, browser, and graphics driver. A repeatable benchmark uses fixed inputs and excludes model downloads and warm-up.
Local artificial intelligence processes data on a personal computer. Language models generate text one token at a time. Prompt processing handles the input before generation begins. Embedding models turn documents into vectors for semantic search. Performance depends on the model, precision, memory, browser, and graphics driver. A repeatable benchmark uses fixed inputs and excludes model downloads and warm-up.
Local artificial intelligence processes data on a personal computer. Language models generate text one token at a time. Prompt processing handles the input before generation begins. Embedding models turn documents into vectors for semantic search. Performance depends on the model, precision, memory, browser, and graphics driver. A repeatable benchmark uses fixed inputs and excludes model downloads and warm-up.
Local artificial intelligence processes data on a personal computer. Language models generate text one token at a time. Prompt processing handles the input before generation begins. Embedding models turn documents into vectors for semantic search. Performance depends on the model, precision, memory, browser, and graphics driver. A repeatable benchmark uses fixed inputs and excludes model downloads and warm-up.
Local artificial intelligence processes data on a personal computer. Language models generate text one token at a time. Prompt processing handles the input before generation begins. Embedding models turn documents into vectors for semantic search. Performance depends on the model, precision, memory, browser, and graphics driver. A repeatable benchmark uses fixed inputs and excludes model downloads and warm-up.
  1. Local language models help people draft and summarize documents without sending their text to a remote server.
  2. Semantic search represents each document as a vector, then finds nearby vectors to retrieve relevant information.
  3. Graphics processors execute many operations in parallel. Memory bandwidth can limit autoregressive token generation.
  4. A reproducible benchmark holds the model, precision, input, and output length constant across measured trials.

Sources