Model Releases

FDTN Antares-1B: Security Vulnerability Detection Agent — A 1B parameter model fine-tuned from Granite-4.0-1B for security vulnerability detection and terminal agent tasks using RL. Small, specialized, and actually useful for security engineering workflows.

OwenSong Inflect-Micro-v2: Edge-Optimized TTS Model — Lightweight VITS-based TTS at 24kHz, optimized for CPU and edge deployment. Apache-2.0 license and PyTorch implementation mean you can actually run this on-device without begging for API credits.

Unsloth Laguna-S-2.1-GGUF: imatrix Quantization for GGUF — Unsloth’s GGUF quantization of Laguna-S-2.1 using imatrix calibration for better quality retention. Comes with conversational format and vLLM compatibility — handy if you’re serving locally.

Open Source Releases

AutoGluon 1.5.1 beta: AutoML framework for tabular, timeseries, multimodal, and text — Amazon’s AutoML library that handles tabular, time-series, multimodal, and text tasks with automatic feature engineering and HPO. Beta release with updated sub-packages. Still AutoML, so your mileage varies, but it’s one of the more battle-tested options.

smcx 1.14.7: State-space inference in JAX with Kalman/particle filters and SMC — JAX-native state-space inference: Kalman filters, particle filters, tempered SMC, and SMC2. Runs on CPU/GPU/TPU with JIT and autodiff. If you’re doing time-series modeling or probabilistic tracking, this saves you from reimplementing filters badly.

vaultspec-rag 0.3.11: GPU-accelerated hybrid dense+sparse RAG with Qdrant backend — Local-first RAG combining dense and sparse retrieval on Qdrant with GPU acceleration. Exposes an MCP-compatible server for Claude Code and other MCP clients. Targets code and document search with hybrid ranking — practical if you’re tired of SaaS RAG pricing.

iterate-ai 0.3.1: Autonomous ML iteration agent with notebook execution — An agent that plans experiments, writes and executes real notebook cells against your data, and ships the best result as a runnable notebook. Supports local (Ollama) and cloud LLM backends with persistent memory. The “bounded autonomy” framing suggests they’ve thought about the failure modes.

pgmnemo-mcp 0.14.1: MCP server for AI agent memory via pgmnemo — MCP server wrapping pgmnemo’s ingest and recall for persistent agent memory. Standardizes memory access across MCP-compatible clients. Useful if you’re building agent workflows that need context to survive session boundaries.

Community Finds

Cursor IDE integrates MCP for persistent coding preferences — Cursor’s MCP implementation for storing and retrieving user coding preferences, project context, and style guides across sessions. The HN thread digs into local-first storage, sync architecture, and privacy tradeoffs — worth reading if you’re evaluating MCP for your own tooling.

Today’s Synthesis

The MCP ecosystem is quietly assembling a practical local-first agent stack. vaultspec-rag gives you hybrid dense+sparse retrieval on Qdrant with GPU acceleration and an MCP-compatible server; pgmnemo-mcp wraps persistent memory ingest/recall behind the same protocol; and Cursor’s MCP integration shows the IDE layer adopting it for cross-session context. Wire these together and you get an agent that retrieves relevant code/docs from your local corpus, remembers decisions across sessions, and surfaces it all inside your editor — no SaaS round-trips, no context window stuffing. Point vaultspec-rag at your repo, run pgmnemo-mcp for the memory layer, configure Cursor (or any MCP client) to talk to both, and you’ve got a coding agent that actually knows your codebase and its history. The HN thread on Cursor’s implementation is worth reading for the sync/privacy architecture details if you’re rolling your own client.