Today's GitHub trending list reads like a quiet reorientation in how developers think about AI: less about scale, more about fit. A 14MB model running on wearables, a graph layer that makes AI reasoning legally defensible, a workspace that gives agents long-term memory across all a team's channels. The common thread is AI that has to earn its place in constrained, accountable, and already-complicated environments.
cactus-compute/needle: A 14MB Foundation Model for Phones and Wearables
Needle 2 is a complete AI model that fits in a single 14MB binary and runs a session in under 28MB of RAM, targeting phones, wearables, smart-home hardware, and robots rather than cloud servers. Despite having only 45 million parameters, it benchmarks competitively against models 5x to 70x larger for the three tasks it's optimized for: tool calling, device control, and structured data extraction. The design choices that make this possible — a Hadamard MLP replacing the standard FFN, grouped-query attention with engram key-value memory, CQ2-bit quantization — are documented in the README and suggest a considered engineering tradeoff rather than a size shrink with capability left on the floor.
github.com/cactus-compute/needle
semantica-agi/semantica: Graph-Native Infrastructure for Accountable AI
Semantica addresses the auditability gap that blocks AI adoption in regulated industries: most agent systems store embeddings, not meaning, leaving no trail a regulator or compliance team can interrogate. The project sits under your existing stack as a deterministic layer that constructs queryable knowledge graphs, runs forward-chaining and Datalog reasoning without touching an LLM, and tags every fact with W3C PROV-O provenance. The recently released v0.6.5 patched five externally-reported security vulnerabilities in the Explorer API and added embedded Oxigraph for RDF storage — a sign that teams are starting to deploy this in production and finding its edges.
github.com/semantica-agi/semantica
macro-inc/macro: One Workspace, One Memory, Every Channel
Macro is an open-source unified workspace that combines email, chat, docs, tasks, calls, agents, and CRM into a single system with shared permissions and context. The differentiator is team-level AI memory: the platform synthesizes conversations, emails, tasks, and completed work into a nightly-refreshed knowledge layer accessible to agents via MCP. Individual tools like Slack and Linear work fine in isolation; the problem Macro is solving is that agents invoked across isolated tools have no coherent picture of what the team actually knows and has decided. The SOC 2 Type II certification and Rust-based CRDT document engine suggest this is aimed at technical teams with real compliance requirements, not just productivity hobbyists.
hugohe3/ppt-master: AI That Generates Real PowerPoint Files, Not Images
PPT Master converts documents and topics into native, editable .pptx files using an AI agent running inside Claude Code or VS Code. The key distinction from screenshot-based approaches is that the output is actual PowerPoint DrawingML: real shapes, connectors, data-backed charts, slide transitions, and audio narration from speaker notes — all locally processed, no platform lock-in. It supports Claude, Kimi K3, GPT, and Gemini with a large context window, and ships as an MIT-licensed skill workflow requiring only Python 3.10+. The project is trending after crossing 46,000 stars, driven partly by a wave of users who tried it for client deliverables and found it substantially faster than building decks by hand.