What happened
has released Decider for Mac, a desktop application that enables users to run the company's 4-billion-parameter decision model, pplx-decider-flash-v1.2, locally on Apple silicon devices. Unlike generative models that produce text, this decision model is designed to analyze text or images and return probabilities for predefined answers, such as yes/no, labels, or numerical scales. The application functions as a local API server, allowing developers to integrate decision-making tasks into local workflows without sending data to external servers.
The Decider for Mac application is a menu bar utility that hosts the pplx-decider-flash-v1.2 model. It requires a Mac with Apple silicon (M1 or later), macOS 13 or newer, 16 GB of RAM, and approximately 6 GB of disk space. Upon the first launch, the app downloads the 5.8 GB model, which may require a Hugging Face token with read permissions.
The application exposes a local API at http://127.0.0.1:8080/v1/decisions, which accepts requests in the same format as 's hosted Decisions API. This allows developers to perform tasks such as binary , multi-label categorization, or scoring on a scale. The app is designed to handle one question per request to optimize performance and accuracy.
provides documentation and Python scripts to demonstrate practical use cases, including sorting files, identifying errors in screenshots, and implementing pre-commit hooks to scan for sensitive information like API keys. These scripts leverage the model's probability outputs to trigger automated actions based on user-defined confidence thresholds.
Source details: pasqualepillitteri.it
Why it matters
Decider for Mac provides a privacy-focused alternative for users who need to process sensitive data—such as private documents, code, or screenshots—without exposing them to cloud-based APIs. By running the model locally, users avoid API costs and data transmission risks. This tool is particularly useful for automated tasks like sorting files, filtering screenshots, or scanning code for secrets before commits, where the model's ability to provide confidence scores allows for threshold-based automation.
The release represents a shift toward localizing specialized AI tasks that do not require generative text capabilities. By focusing on decision-making—assigning probabilities to specific outcomes— addresses a niche for automated data processing where privacy and cost-efficiency are paramount.
The ability to run the model locally allows for the integration of AI into sensitive development pipelines, such as pre-commit hooks, where sending code to a third-party server would be a security violation. This provides a practical, concrete application for local AI that bypasses the latency and privacy concerns associated with cloud-based inference.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
What to watch next
The primary unknowns involve the long-term accessibility of the model and its performance on the minimum hardware requirements. While the model is currently available, the underlying Hugging Face repository is private, requiring a token for access, and it remains unclear if this is a permanent restriction. Additionally, because the model uses a non-standard architecture requiring specific runtime support, users cannot currently load the model into standard third-party local-model runners, limiting its utility to the provided application.
Users should monitor whether opens the model repository or maintains the current private access requirement. The current dependency on a bundled, proprietary runtime prevents the model from being used in standard local-model ecosystems, which may limit its adoption among power users.
Field testing is necessary to determine the model's reliability on the minimum hardware specification (16 GB RAM). Because the model relies on confidence scores, users are advised to validate the model's calibration against their specific datasets before deploying it in automated workflows.