Thinking Machines Lab has released Inkling 2 Mini, a smaller member of its Inkling family of AI models, with its weights open for anyone to download. The model, published on 20 November, has 26.4 billion parameters, only 3.4 billion of which do work on any single token, and is licensed under Apache 2.0.

The company says the model is small enough to run on a single NVIDIA DGX Spark desktop machine or on an Apple silicon computer with at least 32GB of memory.

"Inkling 2 Mini is released with open weights and is available on Tinker today for inference and customization."Thinking Machines Lab, Inkling 2 Mini announcement

The details

  • Design: a mixture-of-experts model, where each token is handled by 8 of 128 specialist sub-networks. Its 24 layers are grouped into four blocks of six, and the two middle blocks are run three times, giving 48 layer passes per token.
  • Looping: each repeat pass adds roughly 1 billion effective active parameters, so with three passes the model works like one with about 5.4 billion active parameters per token.
  • Inputs: it reads text and images and replies in text.
  • Context: up to 1 million tokens, according to the model card.
  • File size: the weights are stored in the compact NVFP4 format (the final four layers in MXFP8), and the checkpoint is about 20GB.

Who gets it and when

The model is available now. The weights are on Hugging Face, and the model can be used and customised through Tinker, the company's own service for running and fine-tuning models. Thinking Machines Lab also says it is offered by outside inference providers, which it did not name in the parts of the page we read. Neither the announcement nor the model card gives a price.

Numbers

The model card compares Inkling 2 Mini with other open models of a similar size, including Qwen3.8-27B. Selected results:

TestWhat it measuresInkling 2 Mini
SWEBench Pro v2 (Hard)Fixing real software bugs77.9%
SWEBench Verified (bash only)Fixing real software bugs79.2%
Terminal Bench v2.1Work in a command line71.6%
BrowseComp (with compaction)Finding facts by browsing the web79.0%
GPQA DiamondGraduate-level science questions86.6%
ARC-AGI-2Abstract puzzle solving49.3%

The table is not one-sided. Qwen3.8-27B scores higher on Terminal Bench (73.0%), GPQA Diamond (90.5%) and the GDPVal-AA work-task test, and the model card lists instruction-following tests MultiChallenge (50.0%) and ComplexConstraints (33.4%) as areas where Inkling 2 Mini trails.

The company also cites an energy-efficiency measure from two Stanford labs, Hazy Research and Scaling Intelligence. On a DGX Spark, Inkling 2 Mini scored 0.703 on "intelligence per watt", against 0.430 for Qwen3.6-35B-A3B and 0.159 for Gemma-4-26B-A4B.

Background

Inkling 2 Mini is part of Thinking Machines Lab's Inkling family of open-weights models. This version is sized to run on a single desktop machine rather than in a data centre. Earlier this month we reported on Tencent's Hy4 model.

What is still unknown

The pages we read do not give a price for using the model on Tinker or name the outside providers that host it. The benchmark scores are the company's own and have not yet been checked by independent testers.

Announced by Thinking Machines Lab on its website: read the announcement

TopicsAI ModelsAI CodingAI AgentsThinking Machines Lab