Model Releases

Ternary-Bonsai-2-27B: Ternary 2-bit Qwen3.8-27B GGUF — Ternary 2-bit quantization of Qwen3.8-27B in GGUF, tuned for llama.cpp with CUDA and Metal. Hybrid attention keeps things interesting; if your deployment target is a phone or laptop, this is worth a look. 🤖

Qwen-2.5-1B-RLCD: MLX structured generation with parallel decoding — MLX build for structured and constrained decoding with parallel decoding on Apple Silicon. Fine-tuned from Qwen2.5-1.5B-Instruct, so it’s small, fast, and aimed at JSON and classification workloads.

Open Source Releases

Recall: Give Claude memory with Redis-backed persistent context — Redis-backed persistent context layer for Claude so your sessions stop having amnesia. Simple API for storing and retrieving conversation state; handy if you’re tired of re-explaining yourself every turn. 🧠

Cloud-Ready Postgres MCP Server — MCP server exposing Postgres to LLMs for natural-language queries and database operations. Built for cloud deployments with secure connection handling, which is a nice change from “works on my laptop” MCP servers.

claude-code v2.1.276 fixes proxy/gateway 400 regression — Hotfix for the 2.1.275 regression where every request died with HTTP 400 “Input tag ‘advisor_20260301’” when ANTHROPIC_BASE_URL points at a proxy or gateway. If you route Claude Code through an API gateway, upgrade before you debug this yourself. 🔥

Research Worth Reading

Position: It is Time to Virtualize Foundation Models with a Self-evolving Operating System Layer — Argues today’s agentic stacks are a mess of implicit runtimes for state, memory, budgets, and guardrails baked into every framework. Proposes virtualizing foundation models behind a self-evolving OS layer so agent behavior becomes portable and governable instead of framework-locked.

MAGS: Multi-agent Auto-formalization Guarantees Safety for Agentic Outputs — Uses multi-agent auto-formalization to apply formal verification to LLM-generated code, catching the safety and security failures fuzzing and static analysis miss. If you’re going to let agents write code, this is the kind of safety net you want.

Closed-World Resolution Against Tool Hallucination in LLM Agents — Identifies a fun failure mode where agents call tools that don’t exist and pass arguments nobody declared — tool selection and gating don’t catch it. The closed-world resolution approach constrains calls to declared schemas, which sounds obvious until you see how often it’s missing.

QVAC Genesis III: A Large-Scale, High-Quality Open Synthetic STEM Corpus for Efficient Language Model Pre-Training — Large-scale synthetic STEM corpus aimed at efficient pre-training for educational models, especially on edge devices with tight token budgets. Addresses the shortage of open, high-per-token-value STEM datasets — useful if you’re training small models on a budget. 📄

Layer-wise Curriculum Learning for Efficient LLM Compression — Transfers knowledge from teacher to student using progressively harder optimization tasks per layer. A solid efficiency play for compressing LLMs without the usual one-shot distillation roulette.

Block Parallelism For Efficient Distributed Long-Context Diffusion Language Model Training — Shards sequences differently than conventional context parallelism to cut distributed attention communication and activation memory overhead. Worth a read if long-context diffusion training is your particular flavor of suffering.

AI Dev Tools

supermemoryai/supermemory: Local memory engine for AI apps — Fast, local memory and context engine with a Memory API for persistent state in agents and assistants. Fully local, so your conversation history doesn’t need to leave the box. 🛠️

ECC: Agent harness performance optimization system — Open-source agent harness focused on performance for AI coding agents, with skills, instincts, memory, and security built in. Works across Claude Code, Codex, OpenCode, Cursor, and friends — think of it as a shared toolbox instead of another silo.

Learn Claude Code: build an agent harness from scratch — Tutorial that builds a nano Claude Code-like agent harness from zero to one in Python. Covers tool use and harness design fundamentals; a solid way to actually understand what your tools are doing under the hood.

JeecgBoot: AI low-code platform with code generation — Enterprise AI low-code platform that generates full-stack Java/Vue3 apps from natural language. Bundles AI chat, knowledge base, workflow orchestration, and MCP plugin support — the kitchen sink, but a well-organized one.

Mcp-Agent – Build effective agents with Model Context Protocol — Framework for building production-grade MCP-based agents with abstractions for tool use, context management, and multi-step reasoning. Aims to cut through the boilerplate so you can get to the actual agent logic.

WhatsApp MCP Server — MCP server that lets LLMs send and receive WhatsApp messages programmatically. A bridge between your agents and the world’s most unhinged group chats. 🫠

Tutorials & Guides

Adventures in Symbolic Algebra with Model Context Protocol — Stephen Diehl connects LLMs to computer algebra systems like SymPy via MCP, showing how it enables reliable symbolic math. Also shares lessons on tool-augmented LLMs that apply well beyond algebra. 📄

Today’s Synthesis

Two things that should never be combined: LLM agents and undeclared tool calls. The closed-world resolution paper identifies a failure mode where agents call tools that don’t exist and pass arguments nobody declared — and tool selection and gating miss it entirely. The fix is mechanical: enforce declared schemas at the boundary. If you’re building with Mcp-Agent , that means a validation middleware that checks every tool call against its declared schema before dispatch — not a prompt instruction, an actual hard check. This matters double when the tool is something like pg-mcp , where a hallucinated argument isn’t a fun bug, it’s a malformed query against your production database. The paper’s point is that this class of error survives selection and gating, so the schema contract has to be enforced in the harness, not the model. Treat MCP’s declared schemas as an API contract, validate every call, and let the agent fail loudly on a 400 instead of silently doing the wrong thing.