Colibri runs frontier MoE models in pure C at 110 stars/day
We don't tell you what's popular — popularity lags and is gameable. We tell you what's gaining speed right now, and whether it's worth your attention. 6 accelerating AI repos earned today's slot.
Top mover
Colibri is a dependency-free C engine that runs large Mixture-of-Experts models by keeping only the router and shared weights in RAM while streaming expert weights from disk on demand, letting a laptop or single workstation serve models far larger than its memory. It's gaining traction because MoE architectures (Mixtral, DeepSeek, Qwen-MoE) make sparse activation practical, and disk-streaming sidesteps the VRAM wall that blocks local inference. The honest caveat: per-token latency is bound by storage throughput, so it trades speed for capacity and is unsuitable for high-QPS serving.
🤖 Agents & automation
Sprix Sage Router implements a state machine that classifies each incoming task as SELF, COLLABORATE, or HANDOFF, letting a network of agents decide locally whether to solve, co-solve, or delegate rather than hardcoding a static topology. It matters now because A2A-style inter-agent protocols are proliferating and most frameworks still route via fixed graphs or LLM-judge prompts, so a lightweight state-aware router is a plausible primitive. Caveat: the repo is young and the 3.6k stars look disproportionate to its maturity, so verify the routing policy, benchmarks, and failure modes before trusting it in production.
growth-lab is an end-to-end agent pipeline that reads your product, pulls data from channels, researches the market, runs campaigns, and feeds results back into the next cycle. It ships as Claude Code / Codex skills, so it slots into existing agent workflows rather than being a standalone SaaS, and its Chinese-market focus (Xiaohongshu, SEO) is a real differentiator. The caveat: 'self-improving growth' depends heavily on the quality of your data connectors and eval loop, which the repo leaves largely to the user.
Tooling & infra
A collection of Claude Code / Codex agent skills that steer LLMs to emit self-contained HTML for wireframes, interactive prototypes, plans, and diagrams instead of Markdown or ASCII. It's gaining traction because HTML artifacts render natively in browsers, support real interactivity and layout, and sidestep the friction of image-based mockups or Mermaid limits. The caveat: it's a prompt/skill pack, not a runtime or library, so output quality still depends heavily on the underlying model and your ability to review generated markup.
Niubigeo runs prompts against LLMs to measure how often a brand appears in AI-generated answers, then produces competitor visibility reports and GEO (generative engine optimization) recommendations. It's gaining traction because 'am I visible in ChatGPT/Perplexity answers' is a real new marketing channel with no standard measurement tooling yet. Caveat: the core method is prompt sampling, so results are noisy, non-deterministic, and easy to game — treat scores as directional, not ground truth.
A curated set of system-prompt rules that instruct coding agents (Cursor, Claude Code, etc.) to avoid telltale AI design patterns: purple gradients, generic hero sections, em-dash-heavy copy, and boilerplate code. It's gaining traction because the bottleneck has shifted from generation to taste — developers are drowning in plausible-but-soulless output and want a cheap, drop-in constraint layer. The caveat: it's a prompt/rules file, not a tool, so effectiveness depends entirely on your agent honoring instructions and your own ability to define what 'not slop' means.