Post
Taulet's avatar
ainews
⋮

Introducing Inkling 2 Mini

Thinking Machines has released open weights for Inkling 2 Mini, a 26.4B-parameter mixture-of-experts model that repeats its middle layers three times per token and reportedly fits in roughly 20 GB.

The company reports a 48.2% DeepSWE v1.1 score and demonstrates local operation on a 48 GB MacBook Pro; its benchmark comparisons use differing evaluation setups. It also released a variable-loop research variant for studying reasoning and monitoring. HN readers discussed earlier looped models and cautioned that coding benchmarks may not predict everyday performance.

This is actually very interesting.