A developer released mlx-decision, a free MIT-licensed Python package that runs open decision models (Cloudflare Clef, Perplexity pplx-decider, d3, H2O-Lightning, LiquidAI d1) locally on Apple Silicon via MLX. It supports fine-tuning, model comparison, and dataset evaluation through a web Studio, Python, CLI, HTTP API, and MCP. Ported models run 1.4–6x faster than the original PyTorch code on MPS.
An experiment demonstrating LoRA fine-tuning of a small vision decision model (d1-2B) on raw game frames so it plays Doom, trained on Apple Silicon in 43 minutes. It relies on mlx-decision, the author's open-source tool for running and training decision models on Macs. The problem solved is showing that game-playing decision models can be trained locally on consumer Apple hardware without large datacenter GPUs.
An experiment demonstrating LoRA fine-tuning of a small vision decision model (d1-2B) on raw game frames so it plays Doom, trained on Apple Silicon in 43 minutes. It relies on mlx-decision, the author's open-source tool for running and training decision models on Macs. The problem solved is showing that game-playing decision models can be trained locally on consumer Apple hardware without large datacenter GPUs.
The model's developer released DecisionTune 1.0, an Apache-2.0 395M encoder (ModernBERT-large plus a 4 KB scoring head) that scores given options rather than generating text, aimed at routine agent decisions like tool routing and yes/no checks. It runs fully locally via PyTorch, MLX, or ONNX backends and ships with an MCP server for local assistants. The developer is seeking feedback on where it picks wrong.
The model's developer released DecisionTune 1.0, an Apache-2.0 395M encoder (ModernBERT-large plus a 4 KB scoring head) that scores given options rather than generating text, aimed at routine agent decisions like tool routing and yes/no checks. It runs fully locally via PyTorch, MLX, or ONNX backends and ships with an MCP server for local assistants. The developer is seeking feedback on where it picks wrong.
kevMac is a lightweight native macOS app that wraps Jared Palmer's open-source kev model, a Jev-compatible decision model built on Qwen. It lets users run Jev-style System One decision models completely locally and offline without writing JSON or Python, returning structured decisions instead of generating text token-by-token.
kevMac is a lightweight native macOS app that wraps Jared Palmer's open-source kev model, a Jev-compatible decision model built on Qwen. It lets users run Jev-style System One decision models completely locally and offline without writing JSON or Python, returning structured decisions instead of generating text token-by-token.
A macOS app that runs small decision-making LLMs locally on the Apple Neural Engine and exposes them as Shortcuts actions. It lets users trigger fast on-device accept/reject-style decisions from OS events like incoming email, file changes, or alarms without writing code. It targets people automating workflows on Apple Silicon who want low-latency, private LLM decisions.
A macOS app that runs small decision-making LLMs locally on the Apple Neural Engine and exposes them as Shortcuts actions. It lets users trigger fast on-device accept/reject-style decisions from OS events like incoming email, file changes, or alarms without writing code. It targets people automating workflows on Apple Silicon who want low-latency, private LLM decisions.
A developer released Snap, a local-first decision engine serving as a drop-in replacement for the TypeSafe System One API, built to test sensitive use cases locally. It ships with the open-weight snap1-2b model under MIT license, classifies, ranks, and decides on inputs in sub-second via single-pass logit reads rather than text generation. Installable on Mac via Homebrew, it targets practitioners wanting local decision models without cloud subscriptions.
A developer released Snap, a local-first decision engine serving as a drop-in replacement for the TypeSafe System One API, built to test sensitive use cases locally. It ships with the open-weight snap1-2b model under MIT license, classifies, ranks, and decides on inputs in sub-second via single-pass logit reads rather than text generation. Installable on Mac via Homebrew, it targets practitioners wanting local decision models without cloud subscriptions.