Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released Inkling, a large open‑weights artificial intelligence model designed to be downloaded, customized and deployed by external developers and companies. The launch marks the lab’s first foundation model trained from scratch and positions Inkling as a general‑purpose system rather than a single flagship competitor to the largest proprietary models.
Inkling uses a Mixture‑of‑Experts transformer architecture with 975 billion total parameters, of which about 41 billion are active for any given task, a design intended to balance scale with efficiency. The model was pretrained on 45 trillion tokens spanning text, images, audio and video, and accepts text, image and audio inputs while producing UTF‑8 text outputs, including code and structured formats. It supports context windows of up to 1 million tokens, putting it among the longer‑context open‑weights systems publicly described so far.
The company is presenting Inkling primarily as a base model for fine‑tuning and domain adaptation rather than an off‑the‑shelf assistant, emphasizing its flexibility and controllable “thinking effort” to trade off latency, cost and answer quality. According to the launch materials, the model is intended for uses ranging from coding assistants and agentic tool‑use systems to general conversational and multimodal applications, with support for multiple programming languages and natural languages.
Inkling’s weights are available under an open‑weights regime, enabling organizations to download and run the model on their own infrastructure, while commercial access is offered through Thinking Machines’ Tinker platform and a network of partner providers. Tinker exposes Inkling for fine‑tuning with 64,000 and 256,000 token context options, and the company is promoting a limited‑time 50% discount for usage on the platform, according to the announcement. Beyond Tinker, Inkling is accessible via APIs from providers including Together AI, Fireworks, Modal, Databricks and Baseten, and the full weights, including an NVFP4 checkpoint for Nvidia Blackwell systems, are hosted on Hugging Face.
The release also introduces Inkling‑Small, a 276‑billion‑parameter Mixture‑of‑Experts model with 12 billion active parameters that is currently in preview. Early materials suggest the smaller variant is intended to provide lower‑cost, lower‑latency performance while matching or exceeding the larger model on selected benchmarks, though detailed comparative results remain limited to the company’s own published metrics.
The launch arrives amid intensified competition in open-weight AI from other major families such as Meta’s Llama and Alibaba’s Qwen, and follows Thinking Machines’ high-profile fundraising and infrastructure partnerships that include a multibillion-dollar Google Cloud deal and backing from investors and hardware partners. How well Inkling performs in enterprise deployments will depend on developer adoption and real-world testing via Tinker and partner integrations.
Read Article: TCS and Google Cloud Launch Gemini Experience Center in Kolkata to Accelerate AI Adoption
