Model Releases

  • tencent/Hy3 — Tencent’s Hunyuan v3 MoE in Apache-2.0 transformers format, conversational and US-region hosted. Another entry in the “we also have a competent open model” sweepstakes 🤖
  • Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF — A 35B-A3B multimodal coder MoE quantized to GGUF, with tool-use, function-calling, and long-context support across five languages. If you run llama.cpp and want a Swiss Army knife that also does vision, this is a dense one to try 🛠️
  • nationaldesignstudio/rampart — A client-side transformers.js PII redaction model built on MiniLM, running entirely in the browser across six languages. Privacy-sensitive text scrubbing without a server round-trip — actually useful for frontend work 🔥

Open Source Releases

  • karakeep-app/karakeep — Self-hostable bookmark-everything app with AI auto-tagging and full-text search across links, notes, and images. For the self-hosting crowd who’ve outgrown their 400 open tabs.
  • steipete/CodexBar — A menu bar tool showing OpenAI Codex and Claude Code usage stats without logging in. Finally, a way to watch your spend anxiety in real time.
  • Daum Search — A search extension pulling from Daum’s web, video, image, blog, book, and cafe verticals, with video and cafe surfaces that differ from Naver. Niche, but if Korean-web coverage matters to you, it’s not redundant.
  • opik 2.1.17 — Comet’s library for logging and evaluating LLM traces. Standard observability plumbing if you’re already in their ecosystem 📄
  • deeppolicy 0.1.0 — An LLM-as-a-judge framework for scoring outputs against policies, rules, and plans. Early release, but beats hand-rolling eval prompts for the tenth time.
  • kelam 0.2.0 — A CLI for building and running voice AI agents straight from the terminal. Voice agent scaffolding without the notebook poetry 🛠️

Today’s Synthesis

The throughline today is local-first inference meeting practical ops: nationaldesignstudio/rampart proves you can run PII redaction client-side with transformers.js, and Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF shows a 35B-A3B multimodal coder runs fine under llama.cpp. Concrete move — wire rampart into your frontend to scrub user text before it ever hits a backend, then hand the cleaned payload to a local Qwopus instance for codegen or tool-use. You keep sensitive data off servers and skip the per-call API bill. For tracking whether that setup actually holds up, opik 2.1.17 gives you trace logging on the local side, and deeppolicy 0.1.0 lets you assert the redaction didn’t leak before the model sees it. Not hype — just a stack you can ship without a vendor lock-in apology.