Model Releases

  • ZAI Org GLM-5.1 — A bilingual (English/Chinese) MoE model under an MIT license. Good if you need scalable performance without the proprietary handcuffs. 🤖
  • OpenBMB VoxCPM2 — Diffusion-based TTS that handles multilingual voice cloning across dozens of languages. Apache 2.0, so it’s actually usable in production.
  • Tencent HY-OmniWeaving — A video diffusion model fine-tuned from HunyuanVideo-1.5 for frame interpolation and weaving. Mostly for researchers who enjoy fighting with video weights.
  • Unsloth Gemma-4-31B-it GGUF — Quantized version of Google’s 31B model for llama.cpp. Keeps the multimodal capabilities but actually fits on your hardware. 🛠️

Open Source Releases

  • Archon — A harness builder for AI coding designed to make LLM output deterministic. Because “it worked once in the chat” isn’t a deployment strategy.
  • quantcpp 0.11.0 — A single-header C++ library that compresses KV caches by up to 7x without sacrificing fp32 precision. A rare win for memory-constrained deployments. 🔥
  • opendataloader-pdf — A PDF parser that tries to preserve semantic layout for ML pipelines. Because parsing PDFs is still a nightmare. 📄
  • Kronos — An open-weight foundation model trained specifically on financial market data. Useful if you’re doing risk modeling and don’t want to prompt-engineer a generalist model.
  • matrix-for-agents 0.6.9.34 — Modular architecture for AI agents with interchangeable skill modules. Basically a plugin system for your LLM workflows.
  • dobbe 0.8.0 — A CLI that coordinates multiple Claude Code agents for PR reviews, vulnerability scanning, and DORA tracking. An attempt to automate the parts of software engineering we all hate.

MCP Servers & Integrations

  • SpecLock — A constraint engine and “patch firewall” for code governance. It adds a layer of typed constraints to keep AI coding assistants from hallucinating your architecture away.
  • ArXiv Scout — MCP server for searching arXiv and extracting PDF text. Saves you from manually downloading 20 papers to find one relevant equation. 📄
  • AgentIndex — A discovery platform indexing 20k+ agents across GitHub and HuggingFace. Like a yellow pages for AI agents.
  • Transcriptor — Fetches transcripts from almost every video platform with a Whisper fallback for audio. Useful for turning video tutorials into LLM context.
  • Technical Analysis — Provides 150+ TA-Lib indicators and stock screening for AI agents. For those building algorithmic trading bots that they hope won’t blow up their accounts.

Tutorials & Guides

  • andrej-karpathy-skills — A distilled CLAUDE.md file based on Karpathy’s observations of LLM coding failures. Practical guidelines to stop Claude from making the same mistakes over and over. 🔥

Today’s Synthesis

If you’re trying to run a large multimodal model on modest hardware without sacrificing determinism, start by pulling the Unsloth Gemma-4-31B-it GGUF quantized checkpoint for llama.cpp, then layer quantcpp 0.11.0 on top to shrink the KV cache by up to 7× while keeping fp32 precision — this cuts memory pressure enough to fit the 31B model on a single 24 GB GPU. Wrap the resulting inference service in a thin agent built with matrix-for-agents 0.6.9.34 so you can hot‑swap skill modules (e.g., a financial‑analysis plug‑in using Kronos or a PDF‑parser using opendataloader-pdf ) without rebuilding the core. Finally, guard the whole pipeline with SpecLock to enforce typed constraints on the LLM’s output, preventing hallucinated architecture changes or unsafe financial advice from slipping into production. The combo gives you a lean, controllable, and extensible AI stack you can actually deploy today.