Thinking Machines Released Inkling 2 Mini, a 26 Billion Parameter Model That Runs Offline on a Laptop
Thinking Machines released Inkling 2 Mini on November 20: a 26.4 billion parameter open-weight model that loops its middle layers and fits in about 20GB.
Thinking Machines just put a coding model on a laptop and told researchers they can watch how it thinks, or miss the part that never gets written down.
The lab released Inkling 2 Mini on November 20: a mixture-of-experts with 26.4 billion total parameters and 3.4 billion active per token. Most of the weights sit in 4-bit NVFP4, the last four layers in MXFP8, and the whole file occupies about 20GB. Open weights, Apache 2.0. The company says researchers ran it offline on a MacBook Pro in airplane mode and had it build a playable flight simulator in a few minutes.
The trick is looping. Inkling 2 Mini has 24 layers. The middle 12 run three times, so each token gets 48 layer applications from 24 stored layers, plus about 5.4 billion weight activations. Thinking Machines says that extra serial depth lets the model reason longer without writing the reasoning down, which is the part a chain-of-thought monitor can actually read.
Small enough to leave the datacenter
The model card dates the release November 20 and lists a 1 million token context window, text and image in, text out. Thinking Machines says Inkling 2 Mini leads every open-weights model under 100 billion total parameters on DeepSWE v1.1 at 48.2 percent of tasks resolved, and on Vals Index v2.2 at 46.5 percent. On the Artificial Analysis Intelligence Index it scores 20.1, tied with Qwen3.8-27B in the table the lab published.
Looping does not store more facts. AA-Omniscience, the lab's factuality score, sits in line with other models of this size. Search and tools close that gap: the same table puts Mini ahead of those peers on BrowseComp, DeepSearchQA, and MCP Atlas.
A research-only sibling, TML-d24-varloop-research, lets testers change the loop count. With no written chain of thought, its time horizon on Thinking Machines' Think Fast bench rose from 16 seconds at one pass to 37 seconds at three. The written traces got shorter as loops increased and, the lab says, stayed legible. Sabotage tests did not trend higher.
Hidden work, smaller file
That is the same fight OpenAI is having in the cloud. Techpresso just covered GPT-6.2 Astra tying protected reasoning to the session that produced it, so a thief cannot lift the hidden chain of thought and train a copy. Inkling 2 Mini is the open-weight version of that tension: more computation per token inside the model, less of it sitting in text someone else can scrape.
The lab is explicit about where Mini is not the best small model. MultiChallenge is 50.0 percent. ComplexConstraints v1.1 is 33.4 percent, behind Qwen3.8-27B on both. Knowledge still scales with parameter count, not with extra loops.
The weights are on Tinker the same day. Hugging Face published a local-setup walkthrough. What Thinking Machines is shipping is not a new chatbot. It is a 20GB file that does more of its thinking before it speaks.
Some offers on this page may be paid placements or contain affiliate links.
Subscribe to Techpresso
Free daily newsletter, read in 5 minutes.
Subscribe free