---
title: "Mira Murati's Thinking Machines ships a 26-billion-parameter model that runs offline on a MacBook"
description: "The 26-billion-parameter open-weights agent shipped hours after OpenAI said attackers had extracted GPT-6 Astra's hidden reasoning at a 54% success rate and asked Congress for new enforcement powers. Together the two moves frame the scarcity bet behind frontier token prices and the market's newest trillion-dollar listing."
publisher: "The Inference"
section: "Model"
published: 2026-11-21T16:11:59.992Z
modified: 2026-11-21T16:11:59.992Z
canonical: https://theinference.org/article/mira-murati-s-thinking-machines-ships-a-26-billion-parameter-model-that-runs-offline-on-a-macbook
language: en
keywords: "AI Models, Open Weights, Thinking Machines Lab, OpenAI, Anthropic"
---

# Mira Murati's Thinking Machines ships a 26-billion-parameter model that runs offline on a MacBook

> The 26-billion-parameter open-weights agent shipped hours after OpenAI said attackers had extracted GPT-6 Astra's hidden reasoning at a 54% success rate and asked Congress for new enforcement powers. Together the two moves frame the scarcity bet behind frontier token prices and the market's newest trillion-dollar listing.

## An agent that works in airplane mode

On November 20, [Thinking Machines Lab](https://theinference.org/markets/companies/thinking-machines-lab), the startup founded by former [OpenAI](https://theinference.org/markets/companies/openai) CTO Mira Murati, [released Inkling 2 Mini](https://thinkingmachines.ai/news/inkling-2-mini/): a 26.4-billion-parameter model whose middle layers run three times per token, so it reasons longer and writes less [1]. It fits in about 20GB, runs on a 32GB Mac, and in the company's demo built a playable flight simulator offline on a MacBook Pro in airplane mode [1]. The weights went to [Hugging Face](https://theinference.org/markets/companies/hugging-face) under an Apache 2.0 license, and the model scores 20.1 on the AA Index, tying [Alibaba](https://theinference.org/markets/companies/alibaba)'s Qwen3.8-27B at roughly a sixth of the cost [1][2].

*Chart: **A laptop model ties the best open peer on the AA Index** Ranked bar chart of AA Intelligence Index v5.0 scores for open-weights models under 100 billion parameters: Inkling 2 Mini 20.1, Qwen3.8-27B 20.1, Nemotron 3.5 Super VL 13.6, Muse Glimmer 6.3, Gemma 4 31B 4.4*

*AA Intelligence Index v5.0 scores, points, for open-weights models under 100B total parameters, as compiled in Thinking Machines Lab's launch post of November 20, 2026. Scores are vendor-reported. [1]*

## Hours earlier, OpenAI counted the theft

[OpenAI's post the same day](https://openai.com/index/combatting-adversarial-distillation) reported that adaptive attackers extracted protected reasoning, the hidden chain-of-thought a model shows no one, from GPT-6 Astra at a 54.0% success rate, and that updated safeguards cut that to 0.2% [3]. Its weekly enforcement against distillation campaigns now ranges from thousands to tens of thousands of accounts [3].

A September disclosure tied a core cluster of one campaign, 16,000 extraction requests from more than 4,000 users in two days, to individuals associated with Moonshot AI, the developer of Kimi [4]. [Anthropic](https://theinference.org/markets/companies/anthropic)'s own [September report](https://techcrunch.com/2026/09/10/anthropic-details-distillation-campaigns-from-alibaba-moonshot-ai-and-deepseek/) counted nearly 200 million exchanges across five campaigns, 151 million of them attributed to Alibaba [5].

*Chart: **OpenAI cut reasoning-theft success from 54% to under 1%** Column chart of reasoning-extraction success rates: against GPT-6 Astra's original safeguards, 54.0% on an earlier request and 22.9% currently; against GPT-6.2 Astra's updated safeguards, 0.2% and 0.9%*

*Success rate of adaptive black-box red-teaming attempts to extract GPT-6 Astra's hidden reasoning, in percent, averaged over six attacker models and measured for an earlier and a current attack request, before and after the safeguards update OpenAI described on November 20, 2026. [3]*

## Whose money the two moves move

Anthropic sold 980 million shares at $83, [raised $81.3 billion in the largest IPO on record](https://www.cnbc.com/2026/11/19/anthropic-ipo-ant-live-updates.html), and closed its debut at $97.17, a $1.8 trillion fully diluted valuation [6]. It fell 8.5% on Friday in a broad tech selloff [7]. Frontier revenue is priced on capability scarcity, and token prices have slid all autumn, as The Inference reported when [OpenAI's newest model ran agent tasks at about a ninth of Claude Opus 5.5's cost](https://theinference.org/article/openai-s-new-model-runs-tasks-about-nine-times-cheaper-than-claude-opus-5-5-the-model-in-anthropic-s-s-1) [7]. The record now shows the teacher leaks while a capable student ships free to anyone with a laptop.

## The fence versus the free copy

OpenAI's counter is that the fence holds and should be law: it backs the Chip Security Act, the AI OVERWATCH Act and the MATCH Act [3]. Murati's caveat cuts both ways. Looping keeps a model's thought process monitorable at this scale, she wrote, and "this may not stay the case" [8]. A model that thinks without writing is harder to audit, whoever downloads it.

## Three numbers decide whether the fence holds

Three numbers decide it: OpenAI's weekly enforcement counts, movement on those bills, and whether open weights keep arriving at frontier-adjacent scores, as [Inkling 2 Small's 594 billion parameters did on November 12](https://thinkingmachines.ai/news/inkling-2-small/) [9]. The teacher's price rests on a fence measured at 0.2%; the student ships free at 26 billion [9].

## Takeaway

A free 26-billion-parameter agent that ties the best open peer shipped the same day OpenAI said reasoning theft from its frontier model fell from 54% to 0.2%.

## Sources

1. [Thinking Machines Lab, "Introducing Inkling 2 Mini," 20 November 2026](https://thinkingmachines.ai/news/inkling-2-mini/)
2. [The Decoder, "New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop," 21 November 2026](https://the-decoder.com/new-open-model-inkling-2-mini-thinks-in-loops-to-solve-problems-with-fewer-words-and-still-runs-on-a-laptop/)
3. [OpenAI, "Combatting adversarial distillation," 20 November 2026](https://openai.com/index/combatting-adversarial-distillation)
4. [Unite.AI, "OpenAI Strengthens GPT-6.2 Astra Against Reasoning Extraction," 20 November 2026](https://www.unite.ai/openai-strengthens-gpt-6-2-astra-against-reasoning-extraction/)
5. [TechCrunch, "Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek," 10 September 2026](https://techcrunch.com/2026/09/10/anthropic-details-distillation-campaigns-from-alibaba-moonshot-ai-and-deepseek/)
6. [CNBC, "Anthropic IPO takeaways: ANT debuts with $1.8 trillion market cap in landmark moment for AI," 19 November 2026](https://www.cnbc.com/2026/11/19/anthropic-ipo-ant-live-updates.html)
7. [Brave New Coin, "Anthropic Stock Price Today: ANT Falls 8% as $500B US Tech Sell-Off Hits AI Stocks," 21 November 2026](https://bravenewcoin.com/insights/anthropic-stock-price-today-ant-falls-8-as-500b-us-tech-sell-off-hits-ai-stocks)
8. [Mira Murati on X, on Inkling 2 Mini and looped-transformer monitorability, 20 November 2026](https://x.com/miramurati/status/2123872534970208421)
9. [Thinking Machines Lab, "Introducing Inkling 2 Small," 5 November 2026 (weights update 12 November 2026)](https://thinkingmachines.ai/news/inkling-2-small/)


---

Mira Murati's Thinking Machines ships a 26-billion-parameter model that runs offline on a MacBook — The Inference. Canonical: https://theinference.org/article/mira-murati-s-thinking-machines-ships-a-26-billion-parameter-model-that-runs-offline-on-a-macbook
