FUTO Swipe

FUTO Swipe is a swipe-typing research release with a dataset, models, and C++ decoding library for turning swipe paths into word predictions.

Research Tools AI Developer Tools Other

Visit Website

Overview

FUTO Swipe is a swipe-typing research and model release focused on turning drawn keyboard paths into word predictions. The project combines a released dataset, trained models, and a decoding library so swipe input can be evaluated and used in inference workflows.

The site describes a model stack with three parts: a universal encoder, a single-language ContextLM, and a layout-specific decoder. Together with dictionary-constrained beam search, these components aim to improve swipe prediction quality while keeping the models small enough to run on low-end devices.

The project grew out of a dataset collection effort on swipe.futo.org in August 2024. That effort produced more than 1 million swipes, and a cleaned dataset of 1 million swipes was released in March 2025 under the MIT license on HuggingFace.

FUTO Swipe also releases swipe-library, a C++ library for inference, decoding, and beam search. The site positions it as the missing piece needed to move from swipe traces to ranked candidate words.

Features

Three-part model stack

Uses three model types together: a layout- and language-agnostic encoder, a single-language ContextLM, and a layout-specific decoder for higher accuracy.

Layout-aware prediction

The encoder provides general swipe prediction across layouts and languages, while the decoder is trained for a specific layout and language to capture keyboard-specific behavior.

Sentence-context filtering

ContextLM uses preceding sentence context to suppress word sequences that do not make sense, and it only needs text data for training.

Beam-search decoding

A dictionary-constrained beam search scores candidate words from swipe paths, which is necessary because raw model outputs are not enough on their own.

C++ inference library

The release includes swipe-library, a C++ library that handles inference, decoding, and beam search from swipe paths to word predictions.

Released swipe dataset

The site reports a released dataset of 1 million English QWERTY swipes under the MIT license, used to train and evaluate the models.

Use Cases

Researching swipe-typing methods
Use the released dataset and model descriptions to study swipe-typing systems, compare approaches, or reproduce evaluation on a controlled English QWERTY corpus.
Building a swipe prediction workflow
Use the encoder, ContextLM, and decoder together when building or testing a swipe-to-word pipeline that needs candidate ranking from raw gesture paths.
Turning swipe paths into ranked words
Use swipe-library when you need the decoding and beam-search layer around model outputs, rather than only a raw model that emits scores.
Comparing model roles and tradeoffs
Use the released resources to understand how much context, layout knowledge, and layout-specific training contribute to prediction quality.
Deploying on constrained hardware
Use the small-footprint model stack when targeting devices or environments where milliseconds-scale inference and low compute use matter.

Pros and Cons

Pros

Releases a large swipe dataset under the MIT license.
Combines layout-agnostic, context-aware, and layout-specific models for prediction quality.
Uses a small parameter footprint that the site says can run on low-end devices in milliseconds.
Includes swipe-library, a C++ component for inference and decoding workflows.
Provides concrete evaluation numbers, including a reported top-4 fail rate of about 4% on the site’s test set.

Cons

The decoder is only available for QWERTY English on the current release.
The source does not describe a hosted app, API, or SDK for direct end-user integration.
Model details and evaluation are specific to the site’s own test set, so broader performance claims are not provided.

FAQ

What does FUTO Swipe do?

The source describes Swipe as a swipe-typing system built from a released swipe dataset, trained models, and a C++ swipe-library for decoding paths into word predictions.

Which keyboard layouts and languages are supported?

The page says the released decoder is currently available only for QWERTY English, because decoder training requires swipe-typing data for a specific layout and language.

Is the dataset open to use?

The site says the dataset was released under the MIT license and is available on HuggingFace.

How does inference work?

The models are described as small enough to run on low-end devices in milliseconds, and the site says the full stack uses a dictionary-constrained beam search to produce candidate words.

Does FUTO Swipe include an API or app?

The page does not list a consumer app, SDK, or hosted API; it presents the models and swipe-library as released resources for swipe-to-word prediction workflows.