FUTO Launches Open-Source Swipe Typing Model to Challenge Big Tech

An Open Challenge to Proprietary Swipe Tech

For years, fast and accurate swipe-typing on mobile devices has been dominated by proprietary, closed-source software from tech giants. These systems often require cloud connectivity, raising privacy concerns for users. FUTO, known for its focus on open-source and user-controlled technology, has now entered the arena with a direct challenge.

The organization has launched FUTO Swipe, a family of open models and algorithms designed to bring high-quality, offline swipe typing to the broader developer community. The primary vehicle is the FUTO Keyboard for Android, but the models are available for anyone to build upon, aiming to break the lock big tech has held on this essential input method.

Building on a Foundation of Real-World Data

Central to FUTO's approach was the creation of a substantial, high-quality training dataset. In August 2024, the company began a voluntary data collection effort via a dedicated website, asking users to swipe sentences, primarily sourced from Wikipedia.

This effort yielded over 1 million individual swipes. After filtering for quality, FUTO released this trove of data under the MIT license on HuggingFace in March 2025. This open dataset not only powers FUTO's own models but also provides a valuable resource for the entire machine learning community working on gesture-based text entry.

A Three-Part Architecture for Precision

FUTO Swipe's performance stems from a clever, multi-model architecture designed to balance accuracy, footprint, and flexibility. The system employs three distinct model types working in concert.

• The Encoder Model: This is a universal, layout-agnostic, and language-agnostic model that handles the core swipe-to-path prediction. While not the most accurate component alone, it provides the foundational understanding of swipe gestures.
• The ContextLM Model: A small, language-specific model that acts as a linguistic sanity check. Trained only on text data, it improves predictions by scoring word likelihood based on preceding context, eliminating nonsensical suggestions.
• The Decoder Model: This is the precision instrument. It is trained on swipe data for a specific language and keyboard layout (currently QWERTY English), learning its unique quirks to deliver leading accuracy.

When combined in a beam search with a width of 300 candidates, this trio achieves a reported top-4 fail rate of only ~4% on FUTO's internal test set. Excluding out-of-vocabulary words, the error rate drops below 1%, a figure the company claims is competitive with established industry leaders.

Surprisingly Light Footprint

Despite its sophistication, FUTO Swipe is remarkably lightweight, a critical feature for on-device, low-latency inference. The encoder contains just 635,140 parameters, and the decoder adds 304,155. The ContextLM is the largest at 1.5 million parameters, but 1.1 million of those are embedding weights.

This brings the total active parameter count to approximately 1.36 million. The entire training process was conducted on a single workstation GPU, highlighting the efficiency of the approach and its low environmental cost compared to massive, cloud-based AI models.

More Than Models: A Complete Inference Library

Recognizing that raw model outputs are not user-ready word predictions, FUTO also releases swipe-library, a C++ library. This handles the entire inference pipeline, including the dictionary-constrained beam search necessary to transform swipe paths into a ranked list of probable words.

This turnkey solution lowers the barrier for developers wanting to integrate swipe typing into novel applications, such as VR interfaces or laptop trackpads, as hinted in FUTO's promotional materials. The models are available under the FUTO Model License (requiring attribution), and the inference library is under GPL.

Context: The Battle for Better Mobile Input

FUTO's release enters a mobile OS landscape where input methods are a key battleground for user loyalty. While not directly about swipe typing, recent coverage of Android and iOS features underscores the constant push for better, more intuitive interaction.

Articles highlight user desire for powerful, yet often buried, features like Android's privacy dashboard, customizable Quick Settings, and advanced multitasking tools—all aimed at making devices more efficient and personal. FUTO Swipe aligns with this trend by offering a core input method that is both powerful and respects user privacy by operating entirely offline.

The development also contrasts with the AI feature wars led by Apple Intelligence and Google's Gemini integrations, proving that significant innovation can still occur in specialized, low-power domains.

The Road Ahead and Why It Matters

FUTO Swipe represents a meaningful step towards democratizing a key mobile technology. By open-sourcing both the models and the training data, FUTO invites scrutiny, improvement, and broader adoption. This challenges the status quo where such capabilities are locked within walled gardens.

For users, the immediate benefit is available in the FUTO Keyboard v0.1.29 for Android, offering a privacy-respecting alternative. For developers, it provides a toolkit to reinvent text entry in new environments. In an era of increasing cloud dependence, FUTO Swipe champions capable, local, and open-source intelligence.