Models
Harness runs a stack of small, specialized models on your device for screen understanding and memory, and routes to frontier language models for reasoning. The on-device models are what let your screen and your memory stay local.
On-device models
These run in your browser, on WebGPU where available and WASM as a fallback, via transformers.js and onnxruntime-web. Nothing they process leaves your machine.
| Role | Model |
|---|---|
| Vision embedding | CLIP ViT-B/16 for scene-change gating and visual search |
| OCR | PaddleOCR PP-OCRv5, with an INT8 path for weaker hardware |
| Dense-text embedding | harness-embedder-v1, fine-tuned by Harness, for semantic screen search |
| Memory embedding | bge-base-en-v1.5 for semantic recall over your memories |
| Retrieval reranker | harness-reranker-v1, fine-tuned by Harness, for reranking recall results |
Fine-tuned models
Harness trained and evaluated its own retrieval models, and they run in the pipeline today:
- harness-embedder-v1 is the dense screen-retrieval embedder, fine-tuned to make semantic screen search sharper than an off-the-shelf embedder.
- harness-reranker-v1 is the episodic-recall reranker, fine-tuned to order recall results by what you actually want back.
Both are published on Hugging Face and run on-device like the rest of the stack.
Language models
For language reasoning, Harness routes to frontier models through providers rather than running the large models on your device. You choose how that inference runs: a local model, your own Venice or Bankr key, or managed inference. The heavier deep-agent reasoning path routes to stronger frontier models when a turn calls for it. See The deep-agent brain and Plans.