Thinking Machines open-sources looped model variant matching mini on coding benchmarks

SonglinYang4 · x · 2026-11-23

Songlin Yang of Thinking Machines released TML-d24-varloop, a research variant (on Hugging Face) of their looped architecture that reportedly matches or beats mini on coding tasks.

The discussion started from @bjoernpl's observation that the model doesn't share KV cache across looped layers, costing 32KB per token vs 16KB with full sharing — meaning far more memory for long contexts than models like Qwen. He suggested a hybrid approach: share the global KV cache while giving each looped layer its own local SWA cache, which reportedly doesn't degrade performance, and asked about ablations since this matters a lot for local inference.

Yang confirmed a terminal KV-sharing variant exists — the newly released model — performing on par with or better than mini on coding.

Original post →

More from Infra

Infra channel →