Thinking Machines Lab has released Inkling 2 Mini, MarkTechPost reports, an open-weight looped transformer that follows Inkling 2 Small. The official announcement describes a Mixture-of-Experts model with 26.4 billion total parameters, 3.4 billion active, and 5.4 billion weight activations per token once looping is counted. The checkpoint is about 20GB, and the lab lists a 32GB Apple silicon Mac, a single NVIDIA DGX Spark, or other GPUs with 32GB or more of memory as supported setups. The weights are on Hugging Face under Apache 2.0, with hosted inference and fine-tuning on Tinker. Ollama, vLLM, llama.cpp, and other runtimes are also listed.
Instead of stacking more unique layers, the model stores 24 layers and runs the two middle blocks three times, for 48 layer applications per token. Thinking Machines reports a 79.2% score on SWE-bench Verified (bash only) and says coding is the model’s strongest area, with weaker results on knowledge recall and some multi-turn instruction-following tests. Those figures are the lab’s own. I have not run Inkling 2 Mini for this post.
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
My take
A 20GB open checkpoint that is aimed at local agentic coding is the part I will actually watch, especially now that Ollama is adding it on Apple silicon. I will treat the benchmark table as the vendor’s own card, and I will try it somewhere I can afford a surprise if I try it at all.
God Bless, and May the Source Be with You!