- Published on
Thinking Machines open sources first multimodal language model, Inkling, focused on low cost and 'resistance to censorship'
- Authors
- Name
- DeepHorus
Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, today released its first major language model, Inkling, under an Apache 2.0 open-source license. The model is designed for enterprises seeking customizable, cost-effective AI that can run on-premises or in virtual private clouds. Inkling boasts a novel "controllable thinking effort" mechanism and a focus on providing factual outputs on sensitive topics, positioning itself as a censorship-resistant option.
The American startup has positioned Inkling as a natively multimodal, Mixture-of-Experts (MoE) system with 975 billion total parameters and 41 billion active parameters. It can reason across text, images, and audio, and its weights are available on Hugging Face and Thinking Machines' own API, Tinker. Alongside the flagship model, the company also previewed Inkling-Small, a 276-billion-parameter variant optimized for low latency and cost.
On third-party benchmarks, Inkling demonstrates strong performance, achieving 77.6% on SWE-bench Verified and 91.4% on VoiceBench. It particularly excels in software engineering and voice understanding, beating US rival Nvidia Nemotron 3. However, Inkling lags behind Chinese open-weight models like GLM 5.2 and DeepSeek V4 Pro on general reasoning and coding tasks. For instance, GLM 5.2 scores 62.1% on SWEBench Pro compared to Inkling's 54.3%.
Inkling's implications for enterprises are significant. Its open-weights nature allows customization and control, while the "controllable thinking effort" mechanism enables balancing performance and cost. The model's designed resistance to censorship may appeal to organizations requiring factual outputs on controversial topics without avoidance.
Industry observers note that Inkling enters a fiercely competitive landscape dominated by highly specialized MoE architectures. Rather than attempting to dominate every benchmark, Thinking Machines has created a broad generalist. The company emphasizes that Inkling's balanced performance across multiple domains makes it a versatile tool for diverse enterprise applications.
Looking ahead, Thinking Machines plans to continue refining Inkling based on community feedback. The availability of open weights and a developer API positions the model for rapid adoption. As enterprises increasingly demand sovereign and customizable AI, Inkling's combination of openness, performance, and cost efficiency could drive significant market impact.
