Skip to header Skip to main navigation Skip to main content Skip to footer
Scott Lawson
Your next upgrade lives here

Thinking Machines Releases Inkling 2 Mini, a Looped Model That Runs on a Laptop

Mira Murati, founder and chief executive of Thinking Machines Lab, photographed at the 2026 Met Gala.

Thinking Machines Lab has released Inkling 2 Mini, an open-weights model designed to run on a laptop rather than in a cloud data center. Released on November 20, the model takes a different approach to reasoning: instead of making the model bigger, it runs part of itself repeatedly.

Inkling 2 Mini is a 24-layer mixture-of-experts transformer with 26.4 billion total parameters and 3.4 billion active parameters on a standard pass. Its first six layers and its final six layers run once for each token, while its middle 12 layers run three times using shared weights. That looping produces 48 effective layer applications per token and 5.4 billion effective active parameters, without increasing the storage size of the model.

The practical result is a checkpoint of about 20 gigabytes that fits on Apple silicon Macs with at least 32 gigabytes of memory, on a single Nvidia DGX Spark, and on other machines with similar memory. Thinking Machines demonstrated the model running offline, in airplane mode, on a MacBook Pro. The model accepts text and images, produces text, and supports a context window of up to one million tokens.

On agentic coding tests, the company and independent evaluators report strong results for a model of this size. Inkling 2 Mini scored 48.2 percent on DeepSWE v1.1 and 79.2 percent on SWE-bench Verified, and it scored 46.5 percent on the Vals Index v2.2. Artificial Analysis gave it a score of 20.1 on its Intelligence Index, tying Alibaba's Qwen3.8-27B, while reporting that Inkling 2 Mini completed tasks roughly four times faster and at about one-sixth the cost per task.

Thinking Machines is led by Mira Murati, the former chief technology officer of OpenAI, who founded the San Francisco lab in early 2025 with other veteran researchers. The company previously released much larger Inkling models, including the 594-billion-parameter Inkling 2 Small earlier in November. Inkling 2 Mini is positioned as the small, local member of that family, aimed at developers and organizations that want a capable coding and tool-use assistant they can run privately on their own hardware.

The lab is also candid about the trade-offs. Because factual recall tends to grow with total parameter count, a compact model cannot match larger systems on every knowledge task. Thinking Machines notes weaker results on some multi-turn chat and instruction-following tests, and occasional repetitive behavior during debugging. Looping also raises an oversight question, since more reasoning happens inside the model and less appears as visible text. Alongside the main release, the company released a variable-loop research version and said it worked with outside groups to study how the internal reasoning can be monitored.

Inkling 2 Mini is available now on Hugging Face under the Apache 2.0 license, with a separate acceptable use policy, and through the company's Tinker platform for inference and fine-tuning. Local runtimes added support quickly after launch, including options for machines with as little as 16 gigabytes of memory using quantized builds. For teams weighing privacy, cost, and control against raw capability, the release offers a clear test: how much useful agent work can now be done without sending every request to the cloud?

Photo: Mira Murati, founder and chief executive of Thinking Machines Lab, at the 2026 Met Gala. Photo by SWinxy, CC BY 4.0, via Wikimedia Commons.

Artificial Intelligence
Technology
Machine Learning
Innovation
Startups
Large Language Models
AI Agents

Copyright © 2026 Rocky Mountain Madman LLC - All rights reserved

Developed and Designed by Rocky Mountain Madman LLC