GitHub Weekly: 10 Rising Stars You Haven’t Heard Of Yet (October 5 – October 11, 2026)

TL;DR: This week we’re skipping the usual overall-trending list and digging into repos that are brand-new (created in the last 30 days) but have already picked up real traction — genuine rising stars rather than the usual mega-projects. The theme that kept showing up: fast, narrow “decision” models as an alternative to full chat LLMs, plus a cluster of tools for running and routing coding agents across different model backends.

Instead of the usual trending-repos format, this week’s GitHub Weekly spotlights 10 rising stars — projects that didn’t exist a month ago but are already earning serious attention. Here’s what stood out for the week of October 5 through October 11, 2026.

#1. storytold/photocraft

github.com/storytold/photocraft — Rust, 49,204 stars

PhotoCraft is a photo editor built entirely from scratch in Rust (a programming language known for speed and memory safety without a garbage collector), rather than wrapping an existing editing engine. It’s part of a small suite of creative tools from the same author — a video editor, a PDF editor, and a vector tool all share the same “from scratch in Rust” philosophy. The appeal is a native, fast alternative to the usual Electron-based or legacy C++ editors.

Good for: Developers interested in how a modern GUI image editor gets built in Rust, and anyone looking for a fast, native photo editor that doesn’t need a subscription or a browser engine running in the background.

#2. NandhaKishorM/laya

github.com/NandhaKishorM/laya — Python, 32,407 stars

Most LLMs generate text one token at a time (“autoregressive”), which is overkill if all you need is a classification, a score, or a yes/no answer. Laya instead makes typed decisions — pick one of several choices, score a piece of text, or answer yes/no — in a single forward pass, across more than 100 languages, using a router that picks the right small model checkpoint for the job. It’s an early entry in what’s becoming a small ecosystem of “decision model” tools (you’ll see the idea come up again later in this list).

Good for: Teams building classification, moderation, or routing logic who want LLM-quality judgment without the latency and cost of a full generative model.

#3. browser-use/jev-ultrafast

github.com/browser-use/jev-ultrafast — Python, 22,661 stars

A “web agent” is an AI system that can navigate real websites — clicking buttons, filling in forms, reading pages — to complete a task on its own. This project, from the team behind the well-known browser-use automation framework, bills itself as the fastest and cheapest option in that space, aimed squarely at the cost and latency problems that make web agents expensive to run at scale.

Good for: Teams building scraping or browser-automation pipelines who need an agent that won’t blow through an API budget on every page load.

#4. Niko1221/Strata

github.com/Niko1221/Strata — C++, 21,952 stars

Strata is a one-click installer and inference engine that runs the open Qwen3.8-Flash-Next model on ordinary consumer Windows or Linux machines, exposing it through an OpenAI/Anthropic-compatible API on localhost (with optional image input). In plain terms: it lets you self-host a capable model and point your existing OpenAI- or Anthropic-flavored code at it without changing much.

Good for: Developers who want a local, free stand-in for cloud LLM APIs during development and testing, especially if their code already talks to OpenAI or Anthropic endpoints.

#5. robbietilton/Compositor

github.com/robbietilton/Compositor — Swift, 16,068 stars

Compositor is a native Mac image editor positioned as a Photoshop alternative — built with Apple’s own Swift language rather than a cross-platform toolkit, which generally means snappier performance and a more “at home on macOS” feel.

Good for: Mac-only designers and developers who want fast, native image editing without an Adobe subscription.

#6. yetone/magpie

github.com/yetone/magpie — Go, 9,774 stars

Magpie is a menu-bar app that lets you point coding agents like Codex or Claude Code at different underlying models — for example, running Codex on DeepSeek or Claude Code on Kimi — all switched from one place instead of reconfiguring each tool separately.

Good for: Developers experimenting with multiple LLM backends for their coding agents who are tired of editing config files every time they want to try a different model.

#7. shihabal3amri/DiPlay

github.com/shihabal3amri/DiPlay — Kotlin, 9,542 stars

CarPlay is Apple’s system for mirroring an iPhone’s interface onto a car’s dashboard screen. DiPlay is an independent, from-scratch implementation of the receiving side of that protocol for Android-based car head units, currently in wired and wireless public preview — letting non-Apple hardware accept CarPlay connections without an official license.

Good for: Hobbyists building or modding car infotainment systems, and owners of Android-based head units who want CarPlay support without buying new hardware.

#8. zai-org/ZCode

github.com/zai-org/ZCode — TypeScript, 7,697 stars

ZCode is a coding-agent harness — the scaffolding that lets an AI model read a codebase, plan changes, run commands, and edit files — built by Z.ai, the team behind the GLM family of open models. It’s a direct competitor to harnesses like Claude Code or Codex, but built to pair naturally with Z.ai’s own models.

Good for: Developers who want an open coding-agent harness built around the GLM model family, or anyone comparing agent harnesses beyond the usual big-name options.

#9. tamaratran/fast-jev-compaction

github.com/tamaratran/fast-jev-compaction — TypeScript, 7,589 stars

When a long Claude Code session runs out of context space, Claude Code “compacts” the conversation history into a summary — which can lose important detail. This plugin replaces that default summary with a scoring pass: every tool call and result gets scored by a fast decision model, low-value ones get dropped or truncated, and everything judged important is kept word-for-word.

Good for: Heavy Claude Code users running long sessions who’ve been burned by the default compaction losing details they needed later.

#10. mizorewww/laya-mlx

github.com/mizorewww/laya-mlx — Python, 6,878 stars

MLX is Apple’s machine-learning framework built specifically for Apple Silicon chips. This project is a native MLX runtime for the Laya decision models from earlier in this list (#2), claiming 7-14ms decisions on an M3 Max with no text generation, no PyTorch, and no cloud API involved.

Good for: Mac developers who want Laya’s fast, typed decision-making running entirely locally and natively on Apple Silicon, with no network round-trip.

Taken together, this week’s rising stars point at two trends worth watching: a growing appetite for small, fast “decision” models as a cheaper alternative to full chat LLMs for classification-style tasks, and an ecosystem of tools springing up to let developers mix and match coding-agent harnesses with whichever model backend is cheapest or fastest that week. None of these are the household names you’d see on an overall-trending list — which is exactly the point of checking in on rising stars.