Model Releases

  • nvidia/Nemotron-Labs-Audex-30B-A3B — A 30B-parameter sparse MoE (3B active) that handles audio understanding, speech recognition, TTS, and text generation in one weights file. If you need a single model for voice-in/voice-out pipelines and you’re US-based, this saves stitching together separate ASR/TTS models — just watch that “other” license. 🤖
  • nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 — A 75B latent-MoE (9B active) quantized to NVFP4 with multi-token prediction, presumably for cheaper inference if you have the right silicon. Loads with custom_code, so you’re running NVIDIA’s Python alongside the weights — fine if you trust them, less so if you audit everything. 📄
  • SupraLabs/Supra-Router-51M — A 51M-parameter Llama-derived router that decides which expert model should handle a prompt, aimed at edge deployments. If you’re building a tiny MoE orchestration layer and want to avoid a cloud call just to pick a model, this fits in a corner of your RAM. 🛠️

Open Source Releases

  • openai/plugins — The official repo for OpenAI’s plugin spec and examples, now publicly hosted. If you have a legacy ChatGPT plugin collecting dust, here’s the source — otherwise it’s a museum piece. 🤖
  • DayuanJiang/next-ai-draw-io — A Next.js app that wires LLMs into draw.io so you can nag your diagrams with natural language instead of dragging boxes. Useful for teams who live in architecture diagrams and would rather type than click. 🛠️
  • gliana-ai — A pay-per-call gateway to 60+ generative models with no API key signup, billing via crypto wallet (MPP/x402). Either a neat way to experiment without vendor onboarding, or a way to lose pocket change to a black box — your call. 🔥
  • socratic-agents 0.4.7 — A multi-agent framework where bots debate under a “constitution” for governance and ethical guardrails. If your agents need a HOA to function, this gives them bylaws. 🛠️
  • ancestryaudit 0.3.2 — A package for spotting and correcting ancestry bias in genomic cancer models, because “one size fits all” rarely fits all genomes. Worth a look if you ship clinical ML and don’t want your ROC curves lying about equity. 📄
  • tpu-inference 0.24.0.dev20260712 — A pre-release Python package for running inference workloads on TPUs, version stamped with a date that implies someone is shipping on a Saturday. If you live in XLA land, maybe pin it; if not, the dev tag says “expect breakage.” 🔥

Today’s Synthesis

If you’re tempted to stand up a local voice agent without babysitting a half-dozen separate services, today’s drops line up neatly: wire SupraLabs/Supra-Router-51M in front of nvidia/Nemotron-Labs-Audex-30B-A3B and nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 to get audio-in/audio-out plus cheaper text generation from one stack. The 51M router decides whether a prompt is speech-driven (send to Audex for ASR + TTS) or pure text (offload to the NVFP4 Puzzle model for multi-token inference on compatible hardware), so you’re not paying active-parameter taxes on every “what’s the weather” query. The neat part is the router fits in a sliver of RAM on the same box — no cloud hop just to decide routing. Just remember Audex sits behind an “other” license and the Puzzle weights need custom_code, meaning your deploy script runs NVIDIA’s Python either way. Pin versions, spin it up in a sandbox, and benchmark real token cost against a single-model baseline before it touches anything prod. 🛠️