The new Inkling 2 Mini model uses a loop‑based transformer architecture to achieve reasoning with fewer tokens, allowing the 26.4 billion‑parameter AI to operate locally on Macs and Nvidia GPUs.

Thinking Machines has unveiled Inkling 2 Mini, a 26.4 billion‑parameter looped transformer that can run on standard laptops equipped with Macs or Nvidia GPUs. The model’s loop‑based architecture reduces the number of tokens needed for reasoning, enabling high‑end performance without cloud resources.

Loop‑Based Architecture Explained

Inkling 2 Mini builds on the original Inkling design by introducing a recurrent loop that re‑processes intermediate representations. This allows the model to reuse its own outputs, effectively compressing the token sequence while preserving context.

The looped approach contrasts with traditional transformers that require a linear increase in tokens for longer inputs. By iterating over a fixed‑size context window, Inkling 2 Mini can handle complex reasoning tasks with fewer computational steps.

Performance on Consumer Hardware

Benchmarks show that Inkling 2 Mini runs comfortably on a MacBook Pro with an M1 Max chip and on laptops featuring Nvidia RTX 3060 GPUs. Users report inference latencies comparable to smaller, non‑looped models, while retaining the expressive power of a 26.4 billion‑parameter network.

Potential Applications

The ability to run a large‑scale model locally opens doors for privacy‑sensitive applications such as on‑device document analysis, offline tutoring assistants, and secure code generation tools.

  • On‑device language translation
  • Secure personal assistants
  • Edge‑based data analytics

For readers interested in broader AI developments, see the AI coverage on uxc.news for related stories and analysis.

The Decoder coverage of Inkling 2 Mini

Track coverage on Momentum · More in AI.