A new open-source project called Ollaya lets users run decision-focused language models on their own machines, eliminating the need for cloud services. The software accepts typed questions about text or JSON and returns calibrated answers in milliseconds, with results processed entirely on local hardware.

Ollaya provides several models optimized for different tasks. The smallest, called decider, operates at 0.75 billion to 1.9 billion parameters and achieves high accuracy on typed decision tasks. Another model, laya, prioritizes speed, delivering responses in under 10 milliseconds for multi-question requests on an NVIDIA RTX 4090 GPU. For safety assessments, qwen3guard evaluates text across 119 languages and categorizes potential risks. The platform also includes von, which handles up to 8,192 tokens of context for complex analyses.

The system uses a single forward pass to generate answers, avoiding the step-by-step token generation typical of conventional language models. This architecture enables rapid responses: a five-question request completes in approximately 178 milliseconds on an RTX 4090, according to internal benchmarks. Median latency measurements from third-party tests show similar performance across supported hardware configurations.

Ollaya's API mirrors TypeSafe's specifications, allowing developers to integrate the models with existing workflows using the official Python SDK. The SDK requires only minimal configuration changes to direct queries to a local Ollaya server. All models are distributed under an Apache-2.0 license, with downloadable binaries for macOS, Windows, Linux, and Docker environments.

Private deployment is emphasized as a core feature. Since processing occurs on user-owned hardware, sensitive data such as internal communications or proprietary documents never leave the local system. The software supports both graphical desktop interfaces and command-line tools across major operating systems, with GPU acceleration available on NVIDIA hardware through WSL 2 or native drivers.

Performance varies by model and hardware. laya demonstrates the fastest response times among the available options, while decider offers the highest accuracy for structured decision tasks. All models operate without requiring internet connectivity once installed, preserving user data within the local environment.