Cactus Hybrid: on-device Gemma 4 that knows when to call the cloud
A 68-param probe scores each answer 0-1, so only 15-35% of queries hit the cloud. Plus a self-hostable Postgres DBA agent and an AI video editor for Claude.
Copy markdown
Local models that admit when they're unsure
Cactus Hybrid bolts a tiny 68-parameter probe onto Gemma 4 E2B that reads hidden states mid-decode and returns a 0-1 confidence for every answer. It scores 0.814 AUROC at spotting its own wrong answers, versus 0.549 for the token-entropy trick most local setups rely on today.
Cut cloud spend by escalating only the hard 15-35%
Route just the low-confidence queries to Gemini 3.1 Flash-Lite and Cactus says you match the big model's benchmarks while keeping 65-85% of traffic on-device. The code is MIT-licensed with copy-paste quickstarts for Transformers, MLX, and llama.cpp, and the routing is a one-line `if confidence < 0.85` check.
DeepSQL: a self-hostable DBA agent for Postgres and MySQL
A new Show HN drops an AI database admin you install in your own VPC in ~15 minutes on read-only replica credentials. It rewrites slow queries, recommends indexes and materialized views, and ships BI dashboards, claiming ~40% lower DB costs, and plugs into Claude, Codex, and Cursor over MCP plus a Slack channel.
Palmier Pro turns your coding agent into a video editor
Palmier Pro is a macOS video editor built to be driven by Claude, Codex, and Cursor over MCP, exposing 400+ agent 'skills' for cuts, effects, and timeline work. It has passed 8,000 GitHub stars, a weekend way to let the agent you already code with edit footage too.