HomeEverythingEducationTV
Equities & FundsCrypto & Digital AssetsAI & TechnologyBusiness & CorporateUS Politics & PolicyGeopolitics & Global RiskMacro, Rates & FXCommodities & EnergyEuropean Politics & MarketsAsia-PacificReal Estate & Property
Story archiveAll categories
← All Stories

Mira Murati Launches Open-Source AI Model Inkling

Created at 16 Jul · 6:31 PM1 source↑ Market-relevant
IN SHORT

Mira Murati's Thinking Machines Lab has released Inkling, a 975-billion-parameter multimodal AI model. Trained from scratch, its full weights are available under an Apache 2.0 license, positioning it as a Western alternative to Chinese models.

✉Newsletter

PiQ Daily

Pick your topics. Get only what matters, on your cadence.

Key Numbers

975 billionInkling total parameters
41 billionInkling active parameters per task
1 millionInkling context window in tokens
45 trilliontokens Inkling was pretrained on
74.1%Inkling MCP Atlas score
30 pointsInkling's lead over Nemotron 3 Ultra on MCP Atlas
77.6%Inkling SWE-Bench Verified score
70.7%Nemotron 3 Ultra SWE-Bench Verified score
82.7%GLM 5.2 Terminal Bench 2.1 score
63.8%Inkling Terminal Bench 2.1 score
78.0%Inkling FORTRESS Adversarial score
276 billionInkling-Small total parameters
12 billionInkling-Small active parameters
$2 billionThinking Machines funding raised in July 2025
$12 billionThinking Machines valuation in July 2025
$50 billionvaluation sought in November 2025 talks

Who's Involved

Mira Murati
Founder of Thinking Machines Lab, formerly CTO of OpenAI
Thinking Machines Lab
AI company that released the Inkling model
Nvidia
Competitor with Nemotron 3 Ultra model
Andreessen Horowitz
Lead investor in Thinking Machines' $2 billion seed round
Z.ai
Developer of Chinese model GLM 5.2
Mira Murati Launches Open-Source AI Model Inkling

↳ Why This Matters

The release of Inkling provides a powerful, fully open-source multimodal AI model from a Western lab, offering an alternative for developers concerned about using Chinese-developed models and potentially fostering greater innovation in the open-source AI community.

Key facts

  • Thinking Machines Lab released Inkling, a 975-billion-parameter multimodal AI model, on July 15.
  • Inkling's full weights are available on Hugging Face under an Apache 2.0 license.
  • Inkling scored 74.1% on MCP Atlas, surpassing Nvidia's Nemotron 3 Ultra.
  • Chinese models like Z.ai's GLM 5.2 and Kimi K2.6 still lead on certain benchmarks.
  • Inkling accepts text, images, and audio inputs with a 1 million token context window.

Mira Murati, formerly the CTO of OpenAI, has launched her new venture, Thinking Machines Lab, with the release of Inkling, a 975-billion-parameter multimodal AI model. The model, trained from scratch, features fully open weights available for download under an Apache 2.0 license on Hugging Face.

Inkling demonstrates strong performance in agentic tasks, scoring 74.1% on MCP Atlas, significantly outperforming Nvidia's Nemotron 3 Ultra. It also achieved a 77.6% score on SWE-Bench Verified. However, Chinese models such as Z.ai's GLM 5.2 and Kimi K2.6 still lead on several key benchmarks, including Terminal Bench 2.1 and Humanity's Last Exam.

Murati left OpenAI in September 2024, approximately 10 months after Sam Altman's brief ousting and reinstatement. She founded Thinking Machines Lab in February 2025 and secured a substantial $2 billion seed round at a $12 billion valuation in July 2025, with participation from major tech firms and venture capital. Earlier reports indicated the company sought a $50 billion valuation in late 2025, but those talks reportedly collapsed by January 2026.

The Inkling model is a mixture-of-experts architecture, designed for efficient inference. It is multimodal, capable of processing text, images, and audio, and supports a large context window of 1 million tokens. The company aims to provide Western developers with a powerful open-source alternative, particularly for those hesitant to use models developed in China due to legal, security, or compliance concerns.

Thinking Machines also previewed Inkling-Small, a smaller model with 276 billion total parameters, which is expected to have its weights released after further testing.

Frequently asked questions

Mira Murati, formerly the CTO of OpenAI, founded Thinking Machines Lab.

Inkling is a 975-billion-parameter multimodal AI model released by Thinking Machines Lab, featuring open weights for download.

Inkling performs well on agentic tasks, outperforming Nvidia's Nemotron 3 Ultra, but Chinese models still lead on some benchmarks.

It provides Western developers with a powerful, transparent alternative to proprietary or Chinese-developed AI models.

What Happens Next

01Weights for Inkling-Small are expected to be released after testing.
02Subsequent fine-tunes of Inkling may be developed to compete with Asian models on specific benchmarks.

Get the newsletter.

Pick the topics you actually care about. We'll email when there's news worth your time, on the cadence you choose. Cancel any time from your account.

Cadence

How It Developed

Mira Murati left OpenAI in September 2024.
Murati founded Thinking Machines Lab in February 2025.
Thinking Machines Lab raised $2 billion at a $12 billion valuation in July 2025.
Talks for a $50 billion raise collapsed by January 2026.
Thinking Machines Lab released Inkling, a 975-billion-parameter multimodal AI model, on July 15.
Inkling's full weights were made available on Hugging Face under an Apache 2.0 license.
Inkling scored 74.1% on MCP Atlas, outperforming Nvidia's Nemotron 3 Ultra.
Inkling scored 77.6% on SWE-Bench Verified.

Sources

T1
Mira Murati Drops Her First AI Model After Leaving OpenAI—And It's Fully Open SourceDecrypt

Related Stories

Moonshot's Kimi 3 AI Model Poised to Rival Anthropic's Opus 4.8
16 Jul · 2:46 PM
Meta Oversight Board: Top AI models avoid criticizing repressive regimes
16 Jul · 10:03 AM
Microsoft trains sales team to criticize rivals' AI models
16 Jul · 12:21 AM
US can maintain AI lead despite China, Anthropic and Pax Silica architect say
16 Jul · 3:06 PM
AMI Labs CEO avoids AGI, superintelligence labels for AI
16 Jul · 3:21 PM