MediaPipe Decision Maker
⚠️ Why not a model
On-device task over EmbeddingGemma, Laya, GLiNER2.5-Decide and Gemma 4. It does not train or ship its own decision weights.
Technical specs
- Base LLMs
- EmbeddingGemma 2, EmbeddingGemma 1, Laya, GLiNER2.5-Decide, Gemma 4 E2B, Gemma 4 E4B
- Decision types
- choice, score, boolean
- Features
- on-device, prewarm, multimodal, jev-json, android, ios, python, web
- License
- Google AI Edge task; each backend keeps its own license
Links
The MediaPipe Decision Maker guide (page marked “Last updated 2026-10-06 UTC”) is an on-device task for structured questions. It returns a winning label, a probability distribution and, for rubrics, an expected score. It does not generate text.
Question types are ChoiceQuestion, BooleanQuestion and ScoreQuestion. evaluateJson accepts a Jev-shaped payload (state plus questions with choice, noul and score) and also maps simple OpenAI JSON Schema fields (enum, boolean, integer range) onto those primitives. Guides exist for Android, iOS, Python and Web.
The documented backends are EmbeddingGemma 1 and 2 (bi-encoder), Laya and GLiNER2.5-Decide (cross-encoder), and Gemma 4 E2B and E4B (single prefill). For embedding backends, prewarm embeds and whitens the options once; later calls score a new context against that cache. The task can share an existing UniversalEmbedder. Inputs can be text, images or 16 kHz audio, including on the options themselves.
Google’s latency and memory lines on that page (“ultra-low latency”, “157 MB bundle”, “0 MB extra RAM” when sharing an embedder) are vendor claims. The task is not a new trained decision model: the weights belong to the backend you load.
Why a runtime entry. Same question as Ollaya and SemIf: no decision-specific checkpoint of its own. A fine-tune such as EmbeddingGemma 2 Typed Decision is a separate catalog model; this task can run the base EmbeddingGemma 2 with zero-shot option matching, which is a different artifact.
