Editorial illustration for Thinking Machines' Inkling 2 Mini: 26B Looped MoE Wins 27 of 28 Tests
Thinking Machines' Inkling 2 Mini: 26B Looped MoE Wins...
Thinking Machines Lab put out a new open-weight model this week, and it skips the usual move of just adding more parameters to get better results. Inkling 2 Mini has 26.4 billion total parameters, but only 3.4 billion are active at a time, and the trick is in how the thing is built. Instead of stacking fresh layers, it reuses the same middle layers three times per token, a method called looping. The payoff is a checkpoint that stays around 20GB, small enough to run on a Mac with 32GB of memory or a single NVIDIA DGX Spark, no data center required.
The model follows Inkling 2 Small, a 594B-parameter release from November 5, and shifts the focus toward something developers can actually run locally for agentic coding, tool use, and document work. It handles text and images as input but only returns text. Thinking Machines shipped it under an Apache 2.0 license on Hugging Face, with hosted inference and fine-tuning available through Tinker, and GGUF builds already up via Unsloth and llama.cpp. Where it lands on benchmarks, and where the looping approach shows its limits, is a different story.
Thinking Machines Lab has released Inkling 2 Mini, an open-weight looped transformer with 26.4B total parameters. The model reuses its middle layers 3 times per token instead of adding more weights. That keeps the checkpoint near 20GB, small enough for a 32GB Mac or 1 NVIDIA DGX Spark.
Why this matters
The 140% inference-compute penalty for the unlooped baseline is the number worth sitting with. Thinking Machines is betting that reusing middle layers three times per token beats the industry's default move of just adding more parameters, and the 27-of-28 eval record plus the 7.5% training-compute tradeoff suggest that bet paid off, at least on 512k-token agentic traces. For developers, the practical pitch is simpler than the architecture: a 26.4B-parameter model that fits on a 32GB Mac or a single DGX Spark, under Apache 2.0, with Tinker handling hosted inference and fine-tuning.
That's a real option for teams who want agentic coding and document tools running locally instead of through an API. For researchers, the 384-experts-with-8-active MoE setup, modified at midtraining, is the part to watch, since it's the lever Thinking Machines pulled to make looping and sparsity work together. We'd want to see how this holds up outside curated evals before calling looped transformers a trend, but as a compute-efficiency argument against brute-force scaling, it's a concrete data point, not a press release.
Further Reading
- Introducing Inkling 2 Mini - Thinking Machines Lab
- Opening the box of Inkling 2 Mini using NNsight+NDIF - NNsight
- Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup - Hugging Face
- Inkling 2 Mini on vLLM: Day-0 Support for a Looped Model - vLLM
- Inkling-2-Mini - NVIDIA