Model Releases

  • LiquidAI/LFM2.5-24B-A2B-GGUF 🤖 — GGUF-quantized variant of LFM2.5-24B-A2B that preserves the MoE sparsity pattern while shrinking the memory footprint enough for on-device inference via llama.cpp. Multi-language conversational AI without giving up the architecture.
  • tencent/Hy-Memory-30B-A3B 📄 — Tencent’s 30B-A3B MoE with Transformers+Safetensors weights, fine-tuned for English and Chinese generation. arXiv reference attached; endpoint-ready if you need a scalable serving setup.

Open Source Releases

  • vLLM: High-Throughput LLM Inference and Serving Engine 🔥 — The default open-serving engine when throughput over novelty matters. Token paging, kernel-level attention optimizations, memory efficiency that actually holds up at scale. Skip it and you’re hand-tuning your own server.
  • parajax 0.4.2 — JAX function vectorization and parallelization you don’t have to hand-tune. If your ML workloads are JAX-based and you’ve been fighting the scheduler manually, this cuts that off at the knee.
  • runledger-ai 0.7.0 — Audit trails, cost tracking, and real-time guardrails for AI coding agents. Generates receipts, risk scores, rollback hooks. Works with Claude Code, Codex CLI, Aider, Factory Droid — basically any agent you’d actually want to constrain.
  • anthropics/claude-code v2.1.304 - Plugin system improvements — Plugin hooks now log which plugin skipped a hook during execution, plus a decider: "person" config that routes permission decisions exclusively to human dialog or host callbacks with refusal reasons. Useful if you actually care about auditability instead of just getting things done.
  • kb-memory 2.1.0 — Local-first agent memory with hybrid retrieval, decay rules, deduplication, and freshness governance. Dual REST + MCP protocols, fully offline, doesn’t need an LLM to function. If you’re privacy-conscious and tired of cloud memory services that phone home, this is the one.
  • parametrix 0.1.10 — Flax-style parameter management for bare JAX. If you miss Flax’s composable abstractions and want them in low-level JAX without the overhead, this gives you that. Fine if you’re already deep in JAX land; skip it if you’re not.

AI Dev Tools

Tutorials & Guides

  • Diagram-Design: Editorial Diagrams for AI Coding Tools — 44 editorial diagram types bundled as self-contained HTML+SVG files, no Mermaid dependency, no shadow rendering issues. If you’re hand-drawing diagrams for Claude Code or Codex prompts and hate Mermaid, this saves you.

Today’s Synthesis

Today’s synthesis connects two governance-focused releases: runledger-ai for runtime audit trails, cost tracking, and rollback hooks on coding agents, paired with kb-memory for local-first agent memory with hybrid retrieval, decay rules, deduplication, and freshness governance. Together they form a self-hosted agent governance stack — replacing cloud memory services and third-party surveillance with full auditability and offline persistence. Add the LiquidAI/LFM2.5-24B-A2B-GGUF quantized variant for on-device inference via llama.cpp, and you get a privacy-compliant coding agent layer running entirely on your own infrastructure. No vendor lock-in, no telemetry, no surprise costs. If your team’s policy requires knowing exactly what your coding agents do at runtime, this stack delivers it without the cloud dependency.