idol — IDOL leverages 3D Gaussian splatting for image-to-3D reconstruction under an Apache-2.0 license, enabling novel view synthesis and 3D content creation from single images with implications for rapid digital twin and asset generation pipelines. 🤖

Pika-H3-Flex — Pika H3-Flex combines diffusers with LoRA adaptation from MiniMax-H3 to support audio-video generation and image-to-video workflows, giving media creators a flexible, framework-integrated pipeline for video generation. 📄

Clover-35B-A3B-Preview — Clover 35B A3B Preview merges mixture-of-experts text generation with diffusion modeling built on Qwen3.6-35B, offering a large-scale conditional diffusion model for research into MoE-driven image generation and editing. 🔧

LFM3-VL-2B — LFM3-VL 2B is a vision-language model from LiquidAI that employs the Liquid architecture for efficient on-device cross-modal processing, supporting multimodal image-text-to-text tasks across 20+ languages with edge conversational AI capabilities. 👁️

Open Source Releases

Kornia: Differentiable Computer Vision Library for Spatial AI — Kornia is a PyTorch-based, differentiable computer vision library offering geometric transformations, augmentation pipelines, and spatial reasoning tools for research and production AI systems, with GPU-accelerated operations for camera pose estimation, stereo matching, and 3D reconstruction. 🛠️

munet-nn 0.4.0 - C++ graph compiler for Vulkan — munet-nn is an experimental C++ graph compiler targeting portable Vulkan for both training and inference, aiming to provide a cross-platform compilation layer that optimizes neural network graphs for Vulkan-enabled hardware, potentially improving performance and hardware accessibility. 🔧

anyray-agent 0.1.4 - Agent reliability transport — anyray-agent provides an httpx2 transport layer that enables OpenAI, Anthropic, and LangChain SDKs to maintain agent continuity when the Anyray gateway is unreachable, implementing intelligent fallback logic that resends calls directly to the provider only when the gateway fails to handle the request. 🔄

orma-ai 0.1.12 - Local episode-first long-term memory — orma-ai introduces a local episode-first approach to long-term memory for small language models, implementing a mechanism for storing and retrieving episodic experiences that enables small LLMs to maintain context and learn from previous interactions without relying on large external memory systems. 🧠

zambo-llamaindex 0.2.6 - Verifiable execution receipts — zambo-llamaindex provides verifiable execution receipts (AER-1 format) for every tool call made by LlamaIndex agents, cryptographically verifying that tool execution occurred as claimed to address transparency and auditability concerns in agent workflows. 🔒

zambo-ai 1.0.11 - Python SDK for verifiable work layer — zambo-ai is the official Python SDK for Zambo, offering a unified MCP endpoint exposing 153 tools via plain JSON-RPC with verifiable receipts for every job execution, operating without API keys or accounts and featuring a free tier of 20 daily calls plus a wrap-up call designed for in-context use. 🆓

AI Dev Tools

AiSOC: Open-Source AI Security Operations Center — AiSOC automates security operations through alert fusion, LLM-agent triage, and MITRE ATT&CK-based investigation workflows, featuring a replayable decision ledger for auditability and full self-hosting without API keys, plus an MCP server compatible with Claude, Cursor, and Continue for integration. 🛡️

AI Engineering Hub: LLM, RAG, and Agent Tutorials — This repository curates in-depth tutorials on LLM fundamentals, RAG architecture design, and real-world AI agent deployment patterns, including code examples, system trade-offs, and performance optimization strategies, serving as a practical guide for building production-grade generative AI applications. 📚

plandex-ai/plandex: Release cli/v2.2.1 — plandex cli/v2.2.1 enables integration with Claude Pro or Max subscriptions for Anthropic model calls, prompting users to connect their subscription when using the default model pack in both cloud and self-hosted modes. 🔗

cc-switch: Cross-platform AI assistant desktop app — cc-switch is a cross-platform desktop application built with Tauri that integrates multiple AI coding assistants including Claude Code, Codex, and DeepSeek Harness into a single workflow, offering WSL support for cross-environment development.

Today’s Synthesis

Combining IDOL’s 3D Gaussian splatting with LFM3-VL-2B’s efficient on-device vision-language architecture creates a powerful local 3D-aware agent. The LFM3-VL-2B provides fast cross-lingual multimodal processing that can understand and reason about visual inputs at edge hardware, while IDOL’s differentiable splatting engine produces high-fidelity 3D reconstructions from single images. Deploying this pipeline via Kornia’s geometric transformation utilities would enable real-time 3D scene understanding and manipulation directly on consumer devices. For production deployment, wrap the workflow in anyray-agent’s reliability transport layer to ensure agent continuity when cloud connectivity is lost, and use zambo-ai’s verifiable execution receipts to audit every reconstruction step. This gives you a self-hosted, auditable 3D-aware agent that can be deployed on-device without compromising safety or performance, while maintaining full compliance with open-weight model licenses.