Thinking Machines shows Inkling, an open source multimodal AI model

Thinking Machines, the AI startup founded by former OpenAI chief technology officer Mira Murati, has released its first major language model under an open source license. Carl Franzen reports for VentureBeat that the model, called Inkling, targets enterprises that want to run AI on their own servers while keeping costs low and avoiding ideological filtering. The company also describes the release in an official blog post.

Inkling is a Mixture-of-Experts system with 975 billion total parameters, of which 41 billion are active for any given response. It processes text, images and audio natively, without bolted-on components for different data types. The model supports a context window of up to one million tokens and was trained on 45 trillion tokens covering text, images, audio and video.

Strong marks, but not the top spot

On benchmark tests, Inkling scores 77.6 percent on SWE-bench Verified, a coding benchmark, beating Nvidia’s Nemotron 3 Ultra at 71.9 percent. It also reaches 91.4 percent on VoiceBench, a voice understanding test, trailing Google’s Gemini 3.1 Pro at 94.4 percent. Chinese models such as GLM 5.2 and Kimi K2.6 outperform Inkling on several coding and reasoning benchmarks. Thinking Machines says it built Inkling as a broad generalist rather than chasing the top score in every category.

A key feature is what the company calls “controllable thinking effort.” Developers can adjust how much computing power the model spends thinking before it answers, on a scale from 0.2 to 0.99. This lets companies save money on simple tasks and increase effort only for complex ones. During training, researchers also observed the model naturally shortening its internal reasoning steps over time without any specific instruction to do so, a phenomenon they call “chain of thought condensation.”

Built to avoid censorship

Thinking Machines states that it trained Inkling to answer directly on politically sensitive topics rather than issuing blanket refusals. An evaluation by AI startup Cognition found that Inkling showed strong resistance to what the company describes as ideological capture. At the same time, the model scored 98.6 percent on StrongREJECT, a test measuring refusal of clearly harmful requests, matching industry safety standards.

The model is released under the Apache 2.0 license, which allows companies to modify and commercialize it without paying royalties. This sets it apart from many other so-called open models that carry usage restrictions. Alongside the main model, Thinking Machines has previewed Inkling-Small, a lighter 276 billion parameter version aimed at applications where speed and cost matter most.

Thinking Machines was founded in late 2024 by Murati along with industry veterans including John Schulman and Barret Zoph. The company raised 2 billion dollars in a seed round led by Andreessen Horowitz in mid-2025, reaching a valuation of 12 billion dollars. Inkling follows the earlier launch of Tinker, the company’s fine-tuning platform, and a research preview of a real-time voice and vision system unveiled in May 2026.

Sources

• Thinking Machines open sources first multimodal language model, Inkling, focused on low cost and ‘resistance to censorship’ – VentureBeat
• Inkling: Our open-weights model – Thinking Machines Lab website