News · 2026-11-21
Thinking Machines releases Inkling 2 Mini with reused transformer layers
Thinking Machines has released Inkling 2 Mini, an open-weight multimodal model that reuses its middle transformer layers three times for each token. The design offers more computation per token without storing a separate set of weights for every pass, and the lab supports local inference on selected machines.
Key facts
- The model was released November 20, 2026, according to Thinking Machines.
- It has 26.4 billion total parameters and 3.4 billion active per pass.
- The official checkpoint has an approximately 20 GB stated memory footprint; the exact download size was not established in the dossier.
- Primary sources: release announcement and model card.
The model has 24 physical layers. Its first six and final six run once, while the middle 12 run three times, producing 48 layer applications per token. Think of a piece of material passing through the same workshop station repeatedly: the factory needs fewer distinct stations, although each pass still takes work. The lab counts about 5.4 billion effective active parameters per token when the repeated passes are included. Its model card calls it a “multimodal autoregressive looped transformer.”
The other efficiency mechanism is a mixture of experts: each token uses a selected subset of the available expert networks. Most weights use a low-precision format called NVFP4; the final four layers use MXFP8. Together these decisions reduce the stored model’s footprint, but repeating computation can affect speed. The looped-transformer lesson and mixture-of-experts lesson unpack those separate choices.
The lab states an approximately 20 GB total memory footprint, which should not be confused with a measured download size. It supports a single NVIDIA DGX Spark and Apple silicon computers with at least 32 GB of memory. The model card recommends 32 GB or more GPU memory for other platforms. Practical headroom can depend on the software and workload. The license is Apache 2.0.
The official release reports 48.2% on DeepSWE v1.1, a coding-agent evaluation, alongside other strong results in its selected comparison. One widely confused pair of numbers illustrates why benchmark details matter: 71.6% is the cited Terminal-Bench v2.1 result using one agent setup, whereas its v4.0 result in another setup is 5.6%. Those are different tests and must not be presented as interchangeable.
Thinking Machines also studied a variable-pass research model. It reports that additional loops increased what models could do without a verbalized chain of thought, while the legibility measures in that study did not meaningfully deteriorate. This is a useful observation at the tested scale, not a guarantee that internal computation will remain observable for every model or task.
Hugging Face’s write-up describes the release, and a separate GGUF conversion provides alternate files for local inference. Those conversions are not the same artifact as the official checkpoint. Community discussion shows interest in local use; it does not independently establish quality, latency or reliability. The practical question is whether the storage savings and repeated computation produce a worthwhile tradeoff on a user’s own hardware and workload.
The arithmetic behind the architecture helps explain why descriptions of the model sometimes use different numbers. Its 26.4 billion total parameters describe stored model capacity. The 3.4 billion figure describes the active parameters in a pass, while the 5.4 billion effective figure accounts for the repeated use of the middle layers. None alone tells a buyer the actual latency, energy use or available context length on a particular machine.
The release also reports safety evaluations. It scores 97.8% on the cited benign FORTRESS test and 89.5% on the adversarial version, where the comparison in the dossier has another model slightly higher. A high score on a named test does not prove resistance to every malicious instruction. The dossier did not find support for a circulating numerical LinuxArena stealth claim, so that number should not be repeated as a result.
An accompanying research variant explores varying the number of repeated passes. The authors compared how much reasoning appeared in the generated text and how legible that text remained under their measures. Shorter visible reasoning does not by itself reveal exactly what happened in hidden layers. Researchers will need other tests before treating the finding as a general rule about whether a model’s internal reasoning can be monitored.
Key questions
Why does Inkling 2 Mini repeat layers?
What hardware does the official model support?
Is 71.6% its Terminal-Bench 4.0 score?
Cite this
APA
Ground Truth. (2026, November 21). Thinking Machines releases Inkling 2 Mini with reused transformer layers. Ground Truth. https://groundtruth.day/news/inkling-2-mini-looped-local-model.html
BibTeX
@misc{groundtruth:inkling-2-mini-looped-local-model,
title = {Thinking Machines releases Inkling 2 Mini with reused transformer layers},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/inkling-2-mini-looped-local-model.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.