Thinking Machines has released Inkling, its first major language model under an enterprise-friendly Apache 2.0 open source license. This model boasts high performance for open weights models on third-party benchmarks, specifically software engineering and voice understanding.
Inkling is a natively multimodal, open-weights Mixture-of-Experts (MoE) system capable of reasoning across text, images, and audio. It has 975 billion total parameters and is designed to balance cost against performance through a novel “controllable thinking effort” mechanism.
The model’s performance is competitive with other open-weight models, including Nvidia Nemotron 3 and DeepSeek V4 Pro. On SWE-bench Verified, it scores 77.6%, beating Nvidia Nemotron 3’s 71.9%. On VoiceBench, it scores 91.4%, compared to Gemini 3.1 Pro’s 94.4%.
Kimi K2.6 outpaces Inkling on several technical benchmarks, but Inkling proves more resilient on general chat instruction following. Against its primary U.S.-based open-weight competition, Inkling demonstrates strong parity and frequent superiority.
The standout feature of Inkling is its “controllable thinking effort.” Developers can programmatically adjust the model’s reasoning budget to dictate how hard the AI should “think” before generating an output. This allows enterprises to deploy Inkling with lower token expenditure for simpler tasks.
In practical terms, this means that enterprises can achieve high-quality results and performance on simple tasks while spending less money. The model’s continuous thinking effort also enables it to reach the same score with a fraction of the tokens, which can be beneficial for AI cost management.
Related: Anthropic Introduces Claude Tag AI Teammate Replacement
A notable element of Thinking Machines’ release is its explicit focus on the model’s epistemics, specifically its calibration, instruction following, and resistance to censorship. Inkling was intentionally trained to answer directly on topics that may be subject to censorship.
Inkling scored 97.1% on AIME 2026 and 77.6% on SWEBench Verified, beating Nemotron’s 94.2% and 70.7%, respectively.
The AI community has praised both the model’s openness and the underlying engineering execution. Thinking Machines co-founder Mira Murati founded the company with the goal of pivoting away from building isolated autonomous agents and towards flexible, multimodal systems.
To understand the significance of Inkling, one has to look back at the rapid trajectory of Thinking Machines. The company’s philosophy began coming into sharper focus with the launch of Tinker, a Python-based API for large language model fine-tuning.
Thinking Machines has given developers a new tool, but it’s also trying to change the way AI development is done. By offering a massive, natively multimodal model under a true open-source license, the company is making a significant contribution to the open-source community.
Inkling’s architecture, specifically its controllable thinking effort, feels like a tangible realization of the company’s goals. With Inkling out in the wild, Thinking Machines has delivered on its foundational promises, attempting to fundamentally rewrite the economics and accessibility of frontier AI development, and it is also available on Hugging Face.
Leave a Reply