Thinking Machines Launches Inkling: A Game Changer for Custom AI Solutions

TL;DR

• **First Open Release:** Thinking Machines, the San Francisco startup founded by former OpenAI CTO Mira Murati, has launched **Inkling**, its first general-purpose open-weight AI model after 18 months of secretive development.
• **Custom AI Hypothesis:** Inkling challenges the "one-size-fits-all" AI paradigm by serving as a foundational framework for organizations to fine-tune tailored solutions using the company's **Tinker** platform, betting that customized models outperform generic ones.
• **Record-Breaking Specs:** The model features **975 billion total parameters** (with 41 billion active per task), supports native text, image, and audio processing, and has debuted as the **highest-scoring open-weight model from a US lab** on the Artificial Analysis Intelligence Index.

Thinking Machines Launches Inkling: A Game Changer for Custom AI Solutions

In a significant move that could reshape the open-source AI landscape, Thinking Machines has officially unveiled **Inkling**, a new open-weight artificial intelligence model designed to disrupt the industry's reliance on universal, one-size-fits-all solutions. Released on Wednesday, July 15, 2026, Inkling marks the first public product from the San Francisco-based startup founded by **Mira Murati**, the former chief technology officer of OpenAI, following a year and a half of development in relative secrecy.

A New Philosophy: Tailored AI Over Universal Models

The core hypothesis driving Inkling is a direct critique of the current AI market. While leading labs like OpenAI, Anthropic, and Google focus on flagship models intended to solve every problem, Thinking Machines argues that AI solutions **tailored and adapted by individual organizations** will ultimately deliver superior performance.

Inkling is not marketed as a final consumer product but rather as a **foundational framework**. The company intends for organizations to download the model's weights and fine-tune them using their own internal data and expertise via **Tinker**, Thinking Machines' dedicated model-customization platform. This approach allows companies to run Inkling locally within their own IT infrastructure, ensuring data privacy and creating highly specific tools for vertical tasks like medical diagnosis or financial risk control.

Technical Specifications: Massive Scale, Efficient Operation

Inkling is architected as a **mixture-of-experts (MoE)** system, boasting a formidable **975 billion total parameters**. However, the model is designed for efficiency: for any specific task, it intelligently activates only a fraction of these—approximately **41 billion parameters**—ensuring faster and more cost-efficient operation.

The model is **natively multimodal**, capable of processing text, images, and audio simultaneously. It was trained on an extensive dataset of **45 trillion tokens** that encompass text, image, audio, and video, enabling it to reason across all these modalities without needing separate adapters. Key technical features include:

• **Context Window:** Supports up to **256K tokens** on the Tinker platform and extends to **1 million tokens** via open weights on Hugging Face.
• **License:** Released under the **Apache 2.0 license**, allowing for commercial use and secondary development.
• **Performance:** Inkling has debuted at **41 on the Artificial Analysis Intelligence Index**, surpassing NVIDIA’s Nemotron 3 Ultra (38) and becoming the highest-scoring open-weight model from a US lab.

Breaking Records and Setting Pricing

The launch has already generated significant industry attention, with Inkling clearing its predecessor's score by three points and putting distance between itself and competitors like Gemma 4 31B and OpenAI’s gpt-oss-120b. On the GDPval-AA v2 benchmark, which scores models on real-world work tasks, Inkling posted an **Elo of 1238**, outperforming Kimi K2.6 and DeepSeek V4 Flash max.

For developers and organizations opting to use Inkling via the Tinker platform rather than running it locally, the pricing is set at **$1.87 per million input tokens** and **$4.68 per million output tokens** at the 64K context tier, with costs roughly doubling at the full 256K window. The model is also noted for its efficiency in coding tasks, matching competitive coding results while burning significantly fewer tokens than rivals.

Availability and Developer Access

Inkling is immediately available through multiple channels. The full open weights are available for download on **Hugging Face**, allowing developers to run the model locally using tools like **llama.cpp** and **Unsloth Studio**. It is also accessible via Thinking Machines’ **Tinker platform** and through various partner integrations.

Soumith Chintala, a prominent figure in the open-source AI community, expressed excitement for the release, highlighting Inkling’s open-weight nature, massive parameter count, and native multimodal capabilities. The company plans to expand the ecosystem further, with announcements expected for a conversational version, **Inkling-Chat**, and a scientific literature-focused variant, **Inkling-Sci**, in the third quarter of 2026.

By releasing Inkling, Thinking Machines is betting that the future of AI lies not in a single super-intelligent model for everyone, but in a flexible infrastructure that empowers organizations to build their own specialized intelligence.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.