Thinking Machines, the American AI startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first major language model under the permissive Apache 2.0 license. With 975 billion total parameters, Inkling is a natively multimodal Mixture-of-Experts system capable of reasoning across text, images, and audio. Designed for enterprises that need customizable, on-premises or virtual private cloud deployment, the model offers strong but not state-of-the-art performance, focusing on adaptability and transparency.
Strong performance in software engineering and voice understanding benchmarks
On SWE-bench Verified, Inkling scored 77.6%, outperforming U.S. rival Nvidia Nemotron 3's 71.9%. In voice understanding, it achieved 91.4% on VoiceBench, close to Gemini 3.1 Pro's 94.4%. However, Chinese models like GLM 5.2 and DeepSeek V4 Pro lead in pure coding and reasoning. Inkling excels in mathematics, posting 97.1% on AIME 2026, and in censorship resistance. The model also surpasses Nemotron 3 Ultra in agentic workflows, with 74.1% on MCP Atlas versus 44.7%.
Sponsored Protocol
Controllable thinking effort for cost-performance trade-offs
Inkling's standout feature is its "controllable thinking effort," allowing developers to programmatically adjust the model's reasoning budget from 0.2 to 0.99. This enables lower token usage for simple tasks and higher compute for complex reasoning. The model's large-scale reinforcement learning over 30 million rollouts led to an emergent phenomenon called "chain of thought condensation," where Inkling learned to compress internal reasoning steps, reducing latency without sacrificing accuracy.
Designed for transparency and resistance to censorship
Thinking Machines explicitly trained Inkling to answer directly on politically sensitive or heavily censored topics, appealing to enterprises that need factual outputs regardless of controversy. In the Propaganda and Censorship Eval, the model showed "strong patterns of censorship non-compliance," avoiding boilerplate refusals. Yet it maintains robust safety: on StrongREJECT, Inkling scored 98.6% against harmful queries. It also achieved a 78.0% refusal rate on adversarial prompts while complying with 95.9% of benign look-alike queries. The company recommends using external moderation tools like Llama Guard for additional security.
Sponsored Protocol
True open-source architecture with Apache 2.0 license
Inkling uses an encoder-free early fusion approach, processing audio as discrete dMel spectrograms and images as 40x40 pixel patches via a hierarchical MLP. With a 1 million token context window and only 41 billion active parameters per token, it is efficient. The Apache 2.0 license allows royalty-free modification, integration, and commercial use. Weights are available on Hugging Face and through the Tinker API. A smaller variant, Inkling-Small with 276 billion parameters, targets low-latency workloads.
Sponsored Protocol
Thinking Machines raised $2 billion in July 2025 and has rapidly built Inkling from scratch. For more on AI security, read our article on OpenAI GPT-Red. For background on open-source licensing, see the Wikipedia page on the Apache License.
