Tenkai Daily — June 20, 2026

Model Releases

• unsloth/GLM-5.2-GGUF — Quantized GGUF version of GLM-5.2 — Unsloth’s GGUF quantization of GLM-5.2 lets you run the MoE model locally with llama.cpp and iMatrix support. Good for edge inference if you don’t want to pull a full FP16 model. 🤖

Open Source Releases

• mnemostack 0.7.0 — Self-hosted hybrid memory & retrieval for AI apps — Hybrid memory stack that fuses semantic, BM25, temporal, and graph recall over Qdrant/Memgraph, with RRF ranking and multi‑tenant filters. Handy if you need a full‑blown retrieval layer baked into your AI app. 🛠️
• tau-coding-agent 0.2.0 — Self-extensible agent CLI with multi-provider LLM support — Terminal UI agent that supports multiple LLM providers, branching sessions, and a plugin system. Great for building a custom coding assistant that can swap backends on the fly. 🐚
• saidso 0.4.1 — Grounding firewall for action-taking AI agents — Forces agents to only use tool arguments explicitly given by the user, with audit logs. Basically a safety net against prompt injection when you let LLMs call real tools. 🛡️
• zlurp: Web Scraping for AI Agents via x402 Micropayments — Turns any URL into clean markdown for agents, paying per‑page via x402 on Base. No API keys, just pay‑as‑you‑go scraping. 4.6k installs say developers actually use it. 📄
• jentic-apitools-score 1.0.0a19 — AI-readiness scoring for OpenAPI specs — Scores OpenAPI specs across six dimensions to see how LLM‑friendly they are. Helpful when you need to gauge whether an existing API is ready for agent tool‑use. 📊
• jentic-apitools-llm 1.0.0a19 — Unified LLM provider abstraction — Provides a common abstraction layer for swapping LLM providers without rewriting code. Reduces lock‑in and simplifies multi‑vendor orchestration. 🔥

Community Finds

• aishwaryanr/awesome-generative-ai-guide — Curated list of GenAI research, interview prep, and notebooks. A one‑stop dump for engineers who want to stay current without scrolling Twitter all day. 📚

Today’s Synthesis

You can bolt a self‑hosted coding assistant together by linking tau-coding-agent 0.2.0 with a local MoE model and a hybrid memory layer. Point the agent’s CLI at the GGUF export of unsloth/GLM-5.2-GGUF so the whole pipeline runs cheap on edge hardware without pulling a full FP16 model. Feed the agent context via mnemostack 0.7.0 , which fuses semantic vectors, BM25 scores, temporal timestamps and graph relationships into a single RRF‑ranked store; multi‑tenant filters let you serve different teams or projects from the same index. The result is a portable, on‑prem coding bot that never leaks prompts to the cloud, can be deployed on a single GPU or a low‑power edge device, and surfaces both recent documentation and legacy code snippets through a unified retrieval layer. You can hot‑swap the LLM backend later via the agent’s plugin system, but the core flow stays unchanged, making the system easy to maintain and scale.