Skip to content
Theme
OnDevice Tools

AIAI Models & Research

Inkling 2 Mini open model reuses layers to reason without extra tokens

Thinking Machines has released Inkling 2 Mini, an open looped transformer that runs half its layers three times per token, cutting output length and cost. It claims top agentic coding scores among open models under 100 billion parameters.

Illustration of a laptop running an open AI model that loops its layers to reason efficiently

Key takeaways

  • Inkling 2 Mini runs 12 of its 24 layers three times per token, reasoning internally without generating extra text.
  • The open mixture-of-experts model has 26.4 billion parameters, 5.4 billion active per token, and an Apache 2.0 licence.
  • It leads open models under 100 billion parameters on agentic coding benchmarks, per The Decoder.
  • It runs locally on Macs with 32 GB of memory or more and on Nvidia's DGX Spark.
In this article
  1. What Thinking Machines released
  2. Size, inputs and licence
  3. Benchmark results and cost
  4. Where it runs and how to try it
  5. Why looped models matter

What Thinking Machines released

Thinking Machines has launched Inkling 2 Mini, an open model built on a looped transformer design. In this approach, the model passes a token through the same layers more than once instead of adding more layers. Inkling 2 Mini runs 12 of its 24 layers three times for each token, according to The Decoder.

The company says this lets the model reason internally for longer without producing extra text. It can reach the right answer using fewer tokens than a conventional model of similar depth.

Size, inputs and licence

Illustration: crypto, AI, code, Hugging Face logo and the labels Hugging Face, Inkling 2 Mini, Looped Transformer

Inkling 2 Mini is a mixture-of-experts model, meaning only part of its network is active for any given token. It has 26.4 billion parameters in total, with 5.4 billion active per token once the loops are counted, Thinking Machines says.

The model handles both text and images and offers a context window of 1 million tokens, which is the amount of text it can consider at once. It is released under the Apache 2.0 licence, so developers can use and modify it freely.

Benchmark results and cost

According to The Decoder, Inkling 2 Mini tops all open models below 100 billion parameters on agentic coding tests, which measure how well a model can carry out multi-step programming tasks.

Artificial Analysis scores it at 20 points on its Intelligence Index. That ties it with Qwen3.8 27B, while Inkling 2 Mini costs roughly six times less and runs about four times faster. These figures come from the companies and benchmark providers cited by The Decoder, not from independent testing by this site.

Where it runs and how to try it

The model is small enough to run locally.

The weights and a research version with variable loops are available for download on Hugging Face, the model-sharing platform. A hosted version can also be tried in the Tinker Playground.

Why looped models matter

Looped transformers are an attempt to get deeper reasoning without paying for a bigger model. By reusing layers, a model can spend more computation on hard tokens and less on easy ones.

If the reported cost and speed figures hold up in practice, models like Inkling 2 Mini could make capable reasoning models cheaper to run on consumer hardware. Readers should treat the benchmark claims as early results until broader independent testing is available.

Sources

  1. New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptopThe Decoder

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.