GenOffice leads as Kimi K3 runs on one CPU
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
GenOffice is a local-first desktop suite that reads and writes real OOXML formats (.docx/.xlsx/.pptx) and does on-device PDF-to-Word conversion, with an AI agent layered on top using your own API key. It matters because it targets the gap between heavyweight LibreOffice and cloud-locked Microsoft 365/Google Workspace, and the BYOK plus local-first model sidesteps per-seat AI pricing. The caveat is fidelity: round-tripping complex Office documents with macros, pivot tables, or advanced formatting is where these projects historically break, so it needs real-world file testing before replacing an existing suite.
🤖 Agents & automation
ARTEMIS is a Google Pixel-Test-Engineering agent that converts natural-language instructions into Android UI automation, capturing logs and exposing an interface for coding assistants like Codex and Claude Code. It reports 99%+ success on the AndroidWorld benchmark, which is notable because reliable mobile GUI agents have lagged behind desktop/web automation. The caveat: benchmark success on AndroidWorld does not guarantee robustness on real apps with dynamic layouts, anti-automation defenses, or non-standard OEM skins.
Firstmate lets you talk to one coordinating agent that spawns and manages a crew of sub-agents to execute tasks, addressing the orchestration gap between single-agent chat and hand-wired multi-agent frameworks. It's gaining traction because developers want delegation without writing graph/DAG plumbing themselves. The caveat: multi-agent coordination adds latency and failure modes, and the repo's maturity and production-readiness are unproven at this star count.
AIPOCH Open-Science is an Electron desktop app bundling scientific agents, Python/R notebooks, data connectors, and MCP support into a model-agnostic, local-first environment for macOS/Windows/Linux. It matters because it targets the reproducibility gap in AI-assisted research by tracking provenance across agent runs and notebook execution, and its 'Claude Science alternative' positioning rides the current wave of interest in agentic scientific tooling. The caveat: at ~4k stars it's still early, and the real value depends on whether its provenance layer is rigorous or just metadata theater.
⚙️ Inference & serving
A from-scratch C99 inference engine that runs a 2.78-trillion-parameter Kimi K3 MoE model on a single CPU within 8.24 GB of RAM, using MXFP4 quantization, AVX2 SIMD, and linear attention to avoid BLAS or any framework dependency. It matters because it demonstrates that extreme-scale MoE inference can be memory-bound rather than GPU-bound, making local/edge deployment of frontier-scale models plausible. The caveat is that CPU-only throughput will be far below GPU serving, so it is best suited for experimentation, constrained environments, or memory-bandwidth-bound batch workloads rather than low-latency production.
Tooling & infra
anti-slop is a set of opinionated Oxlint rules that flag TypeScript/JavaScript patterns lacking type evidence — implicit any, unchecked casts, loose equality, and similar escape hatches that let bugs through. It's gaining traction because Oxlint is fast (Rust-based) and teams are adopting it as an ESLint replacement, so a curated rule pack for strictness fills a real gap. The caveat: it's opinionated by design, so some rules will fight legitimate patterns and require per-rule tuning or overrides.