Agent auto-optimizes new model 7x to 210 tok/s on Apple Silicon with zero guidance
simonguozirui · x · 2026-11-21
Exolabs says it got access to Thinkymachines' new model on Wednesday, hit 30 tok/s within 3 hours, and 7x that (210 tok/s) a day later. This was a real-world test of its upcoming Apple Silicon inference engine and an agent pipeline that continuously optimizes inference across its Mac fleet. Given no guidance about the looping architecture, the agent built a working implementation on its own, discovered architecture-specific optimizations, improved kernels and fusions, and made better use of DSpark via adaptive drafting.
More from coding & agent
- Inkling 2 Mini tested running email, calendar and travel in one open-source agent — stanfordnlp · 2026-11-21
- Stanford's AI Agents course lecture on multi-agent systems praised as most comprehensive — stanfordnlp · 2026-11-21
- Multiplayer coding agents: agent attention is unbounded, human attention isn't — manosaie · 2026-11-21
- Gemini 4 Neon fixes complex codebase issues nearly on par with Opus 5.5 — op7418 · 2026-11-21
- Japanese Devs Say Personal 'Software Garages' Are Ready to Be Shown — 4310sy · 2026-11-21
- What belongs in an agent handoff: evidence, status, and next move — Jdaredevil2 · 2026-11-21