Google DeepMind's case study on Cue is the one piece of outside evidence about a product of ours, so it gets quoted, by us and by others. This post sets out what the page says, so the quotes can be checked against it. The page is at deepmind.google/models/gemma/gemmaverse/cue-ai; its figures are marked current as of May 2026.
What it measured
- The median latency of Cue's dictation polish step fell from 876 ms to 488 ms after the step moved from a cloud model to Gemma 4 E4B running locally through Ollama. The page calls that a 44 percent reduction.
- The measurement was taken on Apple Silicon (M-series) via Ollama, over a benchmark of 227 real voice samples covering English and mixed-language input.
- Dictation per user rose by about 30 percent, measured across active beta users in the four weeks before and after the local model became the default.
- The marginal inference cost of the polish step dropped to zero, which the page gives as the reason dictation is unlimited on every tier, including the free one.
What runs where, according to the page
- The user holds a hotkey and speaks. Audio is transcribed to raw text by a cloud speech-to-text service.
- The raw text is sent to Gemma 4 running locally on the user's machine through Ollama, with a system prompt of about 400 tokens that restores punctuation, segments sentences and removes filler words.
- If Ollama is not running, the desktop app detects that at startup and routes the polish step to a cloud model.
- Cue's agent mode relies on a cloud model. The page reports early evaluations of Gemma 4's function calling for self-contained tasks, and describes them as early.
So “local” on that page means one step of one feature. Transcription is cloud; the agent is cloud; the polish is local by default with a cloud fallback.
How we quote it
The home page uses two figures from it: 488 ms down from 876 ms, and the sentence that Cue is “a voice-activated AI agent that lives on the user's desktop.” Both are on the page in those words. The home page also shows a 45-second segment of a Gemma team talk in which the speaker says engagement rose 30 percent; the written case study says feature usage. We print the written figure and note the difference beside the segment.
What it does not support
- It is not a product review and makes no claim about transcription accuracy or agent task success.
- The 227-sample benchmark it describes is private to Cue and is not published. Our public dataset, the Voice Agent Benchmark Landscape, is a different thing: a map of other people's public benchmarks, with no scores.
- Its figures are dated May 2026. A version of Cue released after that date is not described by the page.
The figures page records how the latency number is used on the home page, and the evidence page lists the case study beside everything else that has been published about our products.