Opening

This week's biggest open-source buzz belongs to Ollaya, which brings the entire pay-as-you-go Jev judgment model stack back to run locally, racking up 509 points and 126 comments on HN. The same week also saw Jared Palmer open-source kev, whose accuracy trails closed-source models by just one percentage point. And if you've been looking for a free Mac alternative to Photoshop, stick around, we'll get to why Compositor hit 5.7k stars in 30 days.

This Week's Open-Source Picks

1. Ollaya: Bringing the Pay-As-You-Go Jev Judgment Model Fully Local

• What it is: Ollaya takes the recently viral, pay-per-use TypeSafe Jev "judgment model" and moves the entire stack to run locally, aimed at engineers who value self-hosting and don't want to send their data out. It's Apache-2.0 licensed and even supports Windows.
• This week's buzz: It topped Hacker News with 509 points and 126 comments, the highest-buzz project of the week.
• Key features: The CLI commands are a direct copy of Ollama's (serve / run / pull / list / ps / show / rm / cp / stop / create), so anyone who's used Ollama has zero learning curve. Its endpoints are wire-identical to TypeSafe (/v1/systemone, /v1/decisions, /v1/models), meaning code that previously hit cloud APIs can run just by swapping the base URL. The binary itself is a compact ~3 MB ONNX graph, while the actual weights are pulled directly from safetensors on Hugging Face, verified with commit pinning and sha256 checks. Official numbers show laya:en answering five questions in just 8 to 10 milliseconds on an RTX 4090.
• Quick start: One line on Linux / macOS with curl -fsSL https://ollaya.dev/install.sh | sh, or irm https://ollaya.dev/install.ps1 | iex on Windows. Once installed, just run ollaya run laya --preset triage "I was charged twice..." to get results.
• Comparison: The key difference from Ollama lies in the model type: Ollama runs generative LLMs that stream tokens one at a time, while Ollaya runs judgment models that return calibrated probabilities in a single forward pass. Compared to TypeSafe's official Jev cloud API, it eliminates the API bill and keeps data on your own machine, though HN commenters have noted it still noticeably lags behind Jev on harder questions.
• Best for: Engineers who value self-hosting and keeping data in-house, and who want to use judgment models for classification, routing, or moderation tasks.
• Link: https://ollaya.dev/

2. kev (jaredpalmer/kev): Train Your Own Jev-Level Judgment Model

• What it is: An open-source judgment model from Jared Palmer, the creator of Turborepo and Formik, built around the idea that "you can train your own version of that closed-source cloud model." The weights are downloadable, and it comes with full training scripts.
• This week's buzz: It shot up to 7.1k GitHub stars in 30 days, with the author's reputation drawing extra attention to this new project.
• Key features: It ships in four sizes ranging from 0.8B to 27B, covering everything from laptops to data centers. A single request can mix true/false, multiple-choice, and rating-scale question types, and every checkpoint comes with temperature calibration for trustworthy confidence scores. The interface aligns with TypeSafe's System One, so existing clients can connect without modification, and you can continue training from an existing checkpoint with a single command. It also includes a Modal deployment script for spinning up an auto-scaling private HTTPS endpoint.
• Quick start: Requires Python 3.12+. Run git clone https://github.com/jaredpalmer/kev.git && cd kev, then uv sync --extra serve, and finally uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009 to spin up a local service. Even the 0.8B version runs fine on an Apple Silicon Mac.
• Comparison: Its direct competitor is TypeSafe's closed-source Jev. The README itself reports Kev-27B scoring 0.848 against Jev's 0.857 on new sources, just a one-point gap, though it still trails on knowledge-heavy tasks (MMLU: Kev-27B at 0.84 vs. Jev at 0.90). Compared with this week's other pick, Ollaya, the two are complementary: Ollaya handles pulling and running the model, while Kev gives you weights you can actually retrain.
• Best for: Teams that want to fine-tune a judgment model on their own labeled data, or need trainable weights instead of a black-box API.
• Link: https://github.com/jaredpalmer/kev

3. Compositor: An MIT-Licensed Native Mac Photoshop Alternative

• What it is: A native Swift image editor for Mac. The author's motivation is refreshingly blunt: Photoshop is too expensive, GIMP feels unnatural to use, so this was built for people who need to edit, remove backgrounds, and design covers, completely free.
• This week's buzz: It hit 5.7k GitHub stars in 30 days.
• Key features: A full layer system with folders, Photoshop-style blend modes, paintable/fillable/blurrable/featherable layer masks, and clipping masks, plus adjustment layers like Hue/Saturation, Curves, Levels, and Gaussian Blur. It can import PSD/PSB files directly and also supports JPEG, PNG, HEIC, TIFF, SVG, and RAW formats. Selection and retouching tools are all here (clone stamp, spot healing brush, magic wand), along with filters for Camera Raw and background removal, all with live preview. It supports non-destructive transforms as well as edge and guide alignment.
• Quick start: Install with one line via brew install --cask robbietilton-compositor, or download it from GitHub Releases. Requires macOS 26.5 or later.
• Comparison: It positions itself squarely as a free alternative to Photoshop, the difference being that Photoshop requires a subscription while Compositor is MIT-licensed and completely free. Compared to GIMP, which is a general-purpose image editor with its own distinct interface conventions, Compositor was designed from the ground up around compositing and post-production workflows, giving it a feel that's closer to what Mac users are already used to.
• Best for: Mac users who want to edit photos, remove backgrounds, and design blog covers for free.
• Link: https://github.com/robbietilton/Compositor

Closing

This week's three tools happen to address two distinct pain points: Ollaya and kev are both about moving pay-as-you-go judgment AI onto your own machine, while Compositor swaps a subscription-based creative tool for a free one. If I could only install one, I'd pick Ollaya: a one-line install lets you take over code that previously hit an API, and you save on the bill immediately. See you next week for more open-source finds.