Thinking Machines releases Inkling, open-source AI model for enterprises

Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, announced the release of Inkling, its first major language model, under an enterprise-friendly Apache 2.0 open-source license. The move targets enterprises seeking to run agentic AI workloads on-premises or in private clouds while retaining full control and customization capability.

Inkling is positioned as a multimodal model with particular strength in software engineering and voice understanding. On third-party benchmarks, it achieves 77.6% on SWE-bench Verified—outperforming fellow open-source competitor Nvidia Nemotron 3, which scored 71.9% on the same test. For voice understanding, Inkling scores 91.4% on VoiceBench, compared to 94.4% for Gemini 3.1 Pro on high reasoning effort. These metrics demonstrate sub-state-of-the-art performance on specialized tasks, positioning Inkling as a credible option for teams prioritizing software engineering or voice applications.

A key differentiator is Inkling's design philosophy around censorship resistance. Thinking Machines notes that the model was built "to answer directly on topics that may be subject to censorship." This design choice directly addresses enterprise concerns about receiving factual outputs irrespective of controversy or sensitivity—a distinction that appeals to regulated industries and organizations where accurate information is critical, regardless of whether the topic touches on contentious areas. The Apache 2.0 license reinforces enterprise appeal by enabling commercial deployment, modification, and redistribution without proprietary restrictions.

Thinking Machines enters the competitive open-source AI market with Inkling at a moment when enterprises are increasingly seeking alternatives to proprietary cloud-based models. The Apache 2.0 license signals a deliberate pivot toward the open-weights segment, where companies like Nvidia have already established footholds with models like Nemotron 3. Mira Murati's presence as founder lends credibility—her OpenAI background suggests familiarity with frontier-model development and enterprise needs.

The benchmarks tell a selective but meaningful story. Inkling's 77.6% on SWE-bench Verified positions it as a credible software engineering assistant, exceeding Nemotron 3 in that domain. However, its 91.4% on voice reasoning trails Gemini 3.1 Pro's 94.4%, indicating it is not universally superior. The multimodal design—spanning software engineering, voice, and text—reflects enterprise demand for all-in-one deployable models. The explicit design choice to answer "directly on topics that may be subject to censorship" addresses a rising concern among regulated industries (finance, healthcare, law) about factual accuracy independent of content-moderation policies. This positioning differentiates Inkling from models optimized primarily for safety-alignment scoring.

No comments yet. Be the first to share your thoughts!