Informe de inteligencia

Thinking Machines' Inkling 2 Mini reuses its own layers to cut token use in an open model

The open mixture-of-experts model loops half its layers three times per token and matches Qwen3.8 27B on one index at a fraction of the cost.

INSIGHTEX Editorial Desk

2 min de lectura

Thinking Machines' Inkling 2 Mini reuses its own layers to cut token use in an open modelInforme de inteligencia

What happened

Inkling 2 Mini, a new open model from Thinking Machines, uses a looped transformer design, The Decoder reported. Rather than sending each token through every layer once, the model passes each token through a set of 12 layers, half of its 24, on three consecutive cycles. The company says this extra internal computation lets the model work through problems without writing out additional text, so it can reach correct answers while using fewer tokens.

Thinking Machines describes Inkling 2 Mini as a mixture-of-experts system with 26.4 billion total parameters, of which 5.4 billion are active per token as a result of the looping. It handles text and images, supports a 1-million-token context window and carries an Apache 2.0 license. Weights and a research variant with an adjustable loop count are on Hugging Face, and the model can be tried in the Tinker Playground.

On agentic coding benchmarks, Inkling 2 Mini posts the highest scores among open models below 100 billion parameters, with a DeepSWE v1.1 result of 48.2 percent and a Vals Index v2.2 result of 46.5 percent. Artificial Analysis rates it at 20 points on its Intelligence Index, level with Qwen3.8 27B, at about one sixth of the cost and roughly four times the speed. It can run locally on Macs with 32 GB of memory or more and on Nvidia's DGX Spark.

Context

Looped transformers pursue stronger reasoning by repeatedly reusing the same layers, spending more computation on each token instead of producing more tokens. Inkling 2 Mini enters a competitive segment of open models under 100 billion parameters, where releases are compared against systems such as Qwen3.8 27B on agentic coding tests and independent rankings like the Artificial Analysis Intelligence Index.

Artificial Analysis profile

Hugging Face profile

FuenteThe Decoder

El boletín

La semana tecnológica, en un solo correo.

Las crónicas que merecen su tiempo, cada semana. Cancele la suscripción cuando quiera.