Google Research's Gemini-SQL2 Dominates Text-to-SQL Benchmarks

Google Research has released Gemini-SQL2, a model that generates SQL queries from natural language. Built on Gemini 3.1 Pro, it recorded **80.04%** accuracy on BIRD, the standard benchmark for the text-to-SQL field — a result that puts it well ahead of the latest models from OpenAI and Anthropic.

Google suggests the technology could contribute to improving natural language interfaces across its data services. It's a development worth watching from the perspective of democratizing data analytics.

Microsoft's SkillOpt: "Training" Markdown Files to Enhance AI Agents

SkillOpt, jointly developed by Microsoft and three Chinese universities, is a method for optimizing instruction documents (Markdown files) for AI agents. By applying the principles of conventional model training, it reportedly **improved GPT-5.5's procedural task performance by roughly 23 points** just by tuning a simple Markdown file.

What makes it particularly interesting is that the same optimized files transfer across different models and agent environments, including Codex and Claude Code. As a new approach that goes beyond the bounds of prompt engineering, its practical impact is substantial.

Pi + Qwen3.6-27B Gains Traction as a Claude Code Replacement

In the local LLM community, a report is drawing attention: the combination of the Pi editor and Qwen3.6-27B is proving to be a practical Claude Code replacement. The reporter has used it daily for over a month, and by pairing it with GPT-5.5 in an advisor role, they were able to almost completely replace Codex and Claude Code.

The Pi editor offers seamless integration of local models, visibility into token usage and costs, and a customizable permissions system, making it an increasingly strong candidate as an open-source local development environment.

llama-launcher v1.3 Auto-Tunes Inference Speed with Bayesian Optimization

Version 1.3 of llama-launcher, a GUI tool for llama.cpp, has been released, adding a Bayesian optimization feature based on Tree-Structured Parzen Estimation. It leverages the optuna framework to automatically adjust llama-server parameters, with reported cases achieving up to a **15% improvement in inference speed**.

It is especially effective for tuning MTP (Multi-Token Prediction) models, making it an attractive tool for local LLM users as a practical alternative to manual parameter tuning.

Using Claude Code as the Commander with Gemini 3.5 Flash as the Implementer

A Qiita post shared a setup guide for using Claude Code as the command center for design and review, while delegating coding implementation to Gemini 3.5 Flash via the Antigravity CLI. It's a useful reference for an approach that achieves both high-quality design judgment and fast implementation while keeping pay-as-you-go costs at zero. The environment was originally built for Claude Fable, but after export regulations made it unavailable, the author switched to Opus 4.8.

---