Technology and Science15 July, 20263 min read
Thinking Machines Lab Releases Inkling Open-Weights Multimodal Model Under Apache 2.0 License
This article is an automated aggregation of coverage from news outlets. Our editorials offer in-depth analysis curated by the NewsCord team.
Inkling is a 975-billion-parameter open-weights multimodal model. Inkling can be downloaded, modified, and run on-premises or in a VPC.
Inkling’s open-weight launch
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first in-house model, Inkling, as an open-weights system under an enterprise-friendly Apache 2.0 open source license. The model is described as a natively multimodal Mixture-of-Experts system with 975 billion total parameters and 41 billion active parameters, and it is built to reason across text, images, and audio. Thinking Machines says Inkling was designed "to answer directly on topics that may be subject to censorship," and it positions the release for enterprises that want to customize, control, and run models on-premises or in virtual private clouds.
The company also previewed Inkling-Small as a lighter 276-billion-parameter alternative optimized for low latency and cost, with weights available on Hugging Face and through its Tinker API. In its own framing, Thinking Machines says Inkling is "not the strongest model available today, closed or open," even as it reports performance on third-party benchmarks such as 77.6% on SWE-bench Verified and 91.4% on VoiceBench.
Benchmarks and competitive field
VentureBeat reports that Inkling scores 77.6% on SWE-bench Verified, beating Nvidia Nemotron 3's 71.9%, and it says Inkling posts 97.1% on AIME 2026 while Nemotron 3 is at 94.2%. The same VentureBeat account places GLM 5.2 ahead of Inkling on SWEBench Pro (Public) with 62.1% versus Inkling’s 54.3%, and it says GLM 5.2 reaches 82.7 on Terminal Bench 2.1 against Inkling’s 63.8. VentureBeat also contrasts Inkling with closed models, saying Claude Fable 5 (max) hits 95.0% on SWEBench Verified and 53.3% on HLE (text only), far outpacing Inkling’s 77.6% and 30.0%.
In a separate framing, TechCrunch reports that Inkling is a mixture-of-experts system with 975 billion total parameters but about 41 billion used per task, and it says the model was trained on 45 trillion tokens of text, image, audio, and video. TechCrunch adds that Thinking Machines pitches Inkling as a starting point for organizations to fine-tune through Tinker, and it notes the company’s claim that Inkling is designed to give calibrated answers, including flagging uncertainty rather than guessing.
Enterprises, Tinker, and stakes
Thinking Machines is positioning Inkling for organizations that want to adapt AI to their own needs, with the model described as downloadable and modifiable directly by outside developers and enterprises rather than accessed only through closed APIs. Bitcoin World says Inkling is trained on 45 trillion tokens spanning text, image, audio, and video, and it describes features including calibrated responses and a user-adjustable ‘thinking effort’ dial that trades depth for speed. The company’s Tinker platform is central to its business model, with Fortune reporting that the startup “generates revenue through its developer tool called Tinker for fine-tuning — or customizing — AI models.”
Fortune also reports that Thinking Machines raised $2 billion at a $12 billion valuation last year, and it says the company is not aiming to monetize the new model itself, instead selling Tinker to customers such as hedge fund Bridgewater Associates. In a concrete example of customization, TechCrunch says researchers from both companies took an existing open-source model and trained it further on Bridgewater’s financial expertise, resulting in a score of 84.7% on financial reasoning tests while costing roughly a fourteenth as much to run, according to the two companies’ own evaluation.