Inkling 2 Mini tops Sonnet 5.5 locally
Thinking Machines Lab's 26B open-weights model clocked up to 210 decode tokens per second on Apple M5 hardware in EXO Labs benchmarks. In local coding evaluations, the looped-transformer model generated a native weather application 1.8x faster than cloud-hosted Claude Sonnet 5.5.
Local inference on unified memory has officially crossed into frontier territory, outpacing premium cloud APIs in interactive developer workflows. By reusing layers in a looped execution cycle, models can achieve deep reasoning while keeping memory requirements low enough for consumer workstations.
- –Blistering Apple Silicon speeds: Reaching 125 tok/s on an M5 Max and 210 tok/s on an M5 Ultra virtually eliminates latency during multi-turn coding loops.
- –Novel looped architecture: Executing 12 middle decoder layers three times provides 48 layers of reasoning depth while keeping memory overhead around 20GB.
- –Real-world agentic speed: Outperforming Claude Sonnet 5.5 by 1.8x on full app scaffolding proves local models can handle real software engineering tasks.
- –Distributed clustering: EXO framework integration allows developers to pool unified memory across multiple Mac devices to run higher precision weights.
DISCOVERED
13m ago
2026-11-21
PUBLISHED
25m ago
2026-11-21
RELEVANCE
AUTHOR
0x0SojalSec