Model Releases

  • Gemma 4 31B Instruct — Google’s 31B‑parameter multimodal model handles image‑text‑to‑text generation, comes with safety fine‑tuning and Apache‑2.0 licensing. Good if you need a conversational vision‑language model that plays nice with standard endpoints. 🤖
  • Gemma 4 26B A4B Instruct — A 26B‑parameter mixture‑of‑experts variant (4 experts) from the Gemma 4 family, also image‑text‑to‑text capable under Apache‑2.0. Handy when you want MoE efficiency without sacrificing multimodal skills. 🛠️
  • LiquidAI LFM2.5 350M — Tiny 350M‑parameter LM built for edge devices, multilingual, Safetensors‑ready, and low‑latency. Ideal for on‑device text generation where memory and power are tight. 📱
  • Ungrounded Gemma 4 26B A4B GGUF — Unsloth’s GGUF‑quantized Gemma‑4‑26B‑A4B‑it for llama.cpp, keeping the original abilities while enabling fast CPU/GPU inference. Includes Imatrix quantization and stays Apache‑2.0. 🚀

Open Source Releases

  • liger-kernel-nightly 0.7.0.dev20260403081233 — Nightly build delivering Triton‑fused kernels for attention and rotary embeddings, aiming to cut memory use and boost training throughput on GPUs. Targets researchers squeezing extra performance from large LLMs. ⚡
  • anthropics/claude-code: v2.1.91 — Adds a _meta['anthropic/maxResultSizeChars'] override to let MCP tools return up to 500K characters, plus a flag to disable inline shell execution in skills and commands. Useful when you need big payloads from Claude‑driven agents. 📄
  • brainctl 0.5.0 — Unified cognitive memory library for AI agents, blending episodic, semantic, and procedural stores with FTS5 text search and vector similarity. Includes neuromodulation hooks and MCP server hooks for distributed agent coordination. 🧠
  • ya-agent-sdk 0.51.3 — Pydantic‑AI‑based framework for building stateful, hierarchical agents with session management, tool abstraction, and logging. Makes modular agent construction less boilerplate‑heavy. 🤖
  • qwen3-embed 1.7.0 — Thin wrapper around Qwen3 text embeddings and rerankers, using ONNX Runtime and GGUF for fast inference. A FastEmbed fork that adds Qwen3 support for low‑latency RAG pipelines. 🔍
  • pdf-autofillr 1.0.7 — Modular PDF form‑filling kit combining a chatbot UI, field mapping, and RAG‑powered completion. Lets you pick and install only the pieces you need for document‑processing workflows. 📑

AI Dev Tools

MCP Servers & Integrations

  • Gmail — Provides end‑to‑end Gmail control for AI agents: send, draft, reply, forward, bulk‑modify/delete, label, archive, trash, and fetch contacts. Mirrors native Gmail capabilities for agent workflows. ✉️
  • Notion — Lets agents search Notion workspaces, view full page details, create/update content, manage databases, and add comments. Brings your knowledge base into agent‑driven processes. 🗂️

Today’s Synthesis

If you’re still routing internal document workflows through paid inference APIs, stop burning budget on token counts and try a fully local stack. Pair LiquidAI LFM2.5 350M with qwen3-embed 1.7.0 and pdf-autofillr 1.0.7 to build a self-contained RAG pipeline that actually respects your data retention policies. The 350M parameter model runs comfortably on consumer edge hardware without melting VRAM, while qwen3-embed’s ONNX/GGUF backend keeps retrieval latency strictly under 10ms. Feed those vectors into pdf-autofillr’s modular field mapper and you get structured extraction without writing brittle regex chains. Wire the components together via a lightweight FastAPI shim, memory-map the quantized weights, and swap out default Python loops for batched CUDA kernels to push throughput past 50 req/s on a single 3090. It won’t replace your 70B fine-tunes for creative generation, but for high-volume form parsing and local compliance QA, it’s the kind of boring infrastructure that actually ships and survives audits. 📉