48 local models tested as agents on M1 Max 64GB: qwen3.8-27b tops real-world score
HAVT_ · reddit · 2026-10-28
A Reddit user built an unattended eval harness and ran 48 local models from their LM Studio library on an M1 Max 64GB as agents: 16 multi-step tool tasks plus 10 tasks rebuilt from real assistant failures, each run 3 times and graded on actual workspace outcomes.
More from coding & agent
- SSH MCP deprecated, succeeded by Rust rewrite SSH-Connect — modelcontextprotocol · 2026-10-28
- adis MCP connector queries Czech VAT unreliability and bank accounts — modelcontextprotocol · 2026-10-28
- Open-source Deepthink claims agent-swarm test-time scaling lifts small-model scores from 14% to 65% — Ryoiki-Tokuiten · 2026-10-28
- Harmonic launches Aristotle: AI that formally verifies code, starting with C and Verilog — wellecks · 2026-10-28
- Dev Calls on X to Ban AI Slop Reply Accounts Flooding viral Threads — generativist · 2026-10-28
- Grok Build v1.0.72 Cuts MCP Context Usage Nearly in Half With Multi-Agent Terminal Upgrades — XFreeze · 2026-10-28