Model Releases

  • prism-ml/Bonsai-27B-mlx-1bit — A 1-bit MLX quantization of Qwen3.6-27B with hybrid attention, built for on-device CUDA/metal inference. If you’re running LLMs on edge hardware and tired of 4-bit being “good enough,” this is worth a look 🤖
  • Wan-AI/Wan-Dancer-14B — Image-to-video model that turns a still image plus music into a dancing sequence, with an Apache-2.0 license and multilingual support. The “music-to-dance” framing is either a party trick or a pipeline component depending on who you ask 📄
  • GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking-GGUF — A 1B GGUF build with thinking, tool-calling, and function-calling wired in for llama.cpp. The naming convention is doing a lot of heavy lifting, but a 1B model that does tool calls locally isn’t nothing 🤖

Open Source Releases

  • anthropics/claude-code: v2.1.214 — Patch release fixing dir/** allow rules that were auto-approving writes outside the intended scope, plus a permission-check bypass fix. Standard security-hole-plugging that you’ll appreciate if you’d been wondering why your guardrails felt porous 🛠️
  • tpu-inference 0.24.0.dev20260718 — New dev prerelease for TPU inference tooling. Pre-1.0 and dated, so treat it as a moving target unless you’re already in that ecosystem.
  • nemo-fabric-adapters-deepagents 0.1.0a20260718 — A LangChain Deep Agents adapter for NeMo Fabric, still alpha. Relevant if you’re stitching agent frameworks onto NVIDIA’s orchestration layer.
  • mcp-embedding-daemon 1.20.0 — A shared ONNX/GGUF embedding server for the n24q02m MCP ecosystem. If your MCP setup is choking on embedding calls, a dedicated daemon beats reloading models per process 🛠️
  • clauderizer 1.11.0 — MCP-native working memory for agents: gameplans, phases, dependency graphs, all over plain markdown. The “post-hoc cascade” bit suggests it tracks how changes ripple — useful if your agents keep forgetting what they were doing.
  • boost-skill-cli 1.0.22 — “Homebrew for AI coding skills” — a CLI for installing and managing reusable agent skills. Catchy analogy, though whether the package manager space needed another entrant is a question for your free time 🛠️

AI Dev Tools

  • tirth8205/code-review-graph — Local-first code intelligence graph for MCP and CLI that maps your codebase so AI tools read only relevant files. Benchmarked context reductions on reviews — if your coding agent keeps drowning in irrelevant files, this is the kind of plumbing that helps 🔥

Today’s Synthesis

The throughline today is local-first agent infrastructure getting quietly serious. tirth8205/code-review-graph cuts agent context bloat by mapping your codebase so tools read only what matters, and mcp-embedding-daemon stops you from reloading embedding models per process by serving them shared over ONNX/GGUF. Pair that with clauderizer , which gives MCP agents persistent working memory via plain markdown, and you’ve got a coherent stack: scope the code, embed it once, and stop your agents from forgetting mid-task. Concrete move — if you’re already on MCP, drop in the embedding daemon and the review graph before your next agent eval; the context reduction is measurable and the memory layer is cheap to trial. The model releases (prism-ml/Bonsai-27B-mlx-1bit , GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking-GGUF ) sit fine on top of that once the plumbing holds.