Summary
- On July 15, Thinking Machines Lab unveiled Inkling, a multimodal AI model with 975 billion parameters, fully trained from scratch and available on Hugging Face under an Apache 2.0 license.
- Inkling achieved a score of 74.1% on MCP Atlas, significantly outperforming Nvidia's Nemotron 3 Ultra by almost 30 points, making it the top-performing open-weights model in the West for agentic tool use. However, Chinese models GLM 5.2 and Kimi K2.6 still excel in several critical benchmarks.
- In July 2025, Thinking Machines secured $2 billion in funding at a $12 billion valuation, later attempting to raise $50 billion in November, though those negotiations ultimately failed by January 2026.
Mira Murati departed from OpenAI in September 2024 to pursue her own ventures. Nearly two years later, she introduced Inkling through her startup, Thinking Machines Lab. This multimodal AI model was developed entirely from scratch, with all weights accessible for free download.
Following the firing of Sam Altman from OpenAI in November 2023, Murati, who was the CTO at the time, briefly became interim CEO before Altman was reinstated five days later. Murati resumed her role as CTO but left OpenAI approximately 10 months later to establish Thinking Machines Lab in February 2025.
After its inception, the company quickly gained significant financial backing, raising $2 billion at a valuation of $12 billion in July 2025, led by Andreessen Horowitz along with Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street—marking one of the largest seed funding rounds in Silicon Valley's history at the time.
By November 2025, rumors indicated the company was looking to raise funds at a $50 billion valuation, but these discussions collapsed by January 2026.
About Inkling
Inkling employs a mixture-of-experts architecture, activating only a portion of the network for each input, ensuring fast inference while maintaining depth. With a staggering 975 billion total parameters and 41 billion active parameters per task, it is impractical to run on personal machines.
This model is multimodal, capable of processing text, images, and audio, and features a context window allowing it to reason over 1 million tokens, roughly equivalent to 750,000 words. It was pretrained on an extensive dataset of 45 trillion tokens, covering text, images, audio, and video.
"Introducing our first model, Inkling. Trained from scratch with open weights, fine-tunable on Tinker today," Murati announced on X. The significance of it being trained from scratch is substantial, especially for the open-source community, as it offers a viable alternative for Western developers who may be hesitant to use Chinese models due to legal or ethical concerns, especially since leading AI firms in the West primarily focus on closed-source models.
Fine-tuning allows for the retraining of an existing model on specialized datasets to enhance its performance for specific tasks. Tinker serves as Thinking Machines' cloud platform designed for this purpose. The complete model weights are also available on Hugging Face under an Apache 2.0 license, with no restrictions.
Inkling excels in agentic tasks. On MCP Atlas, which assesses the reliability of AI agents in completing real-world tasks using Model Context Protocol, Inkling scored 74.1%, significantly surpassing Nvidia's Nemotron 3 Ultra, the primary competitor in the Western open-weights space.
In SWE-Bench Verified, which tests an AI agent's ability to autonomously resolve software bugs on GitHub, Inkling achieved a score of 77.6%, again outperforming Nemotron's 70.7%.
Thinking Machines markets this model as “well-rounded” and generalist, indicating it does not sacrifice quality in specific tasks for the sake of performance in others (unlike models excelling in coding but lacking in creative writing, for instance).
Source: Thinking MachinesChinese models still maintain advantages in several areas. For example, Z.ai's GLM 5.2 scored 82.7% on Terminal Bench 2.1, which evaluates autonomous AI coding agents in real terminal environments, while Inkling managed only 63.8%. Kimi K2.6 outperformed Inkling in Humanity's Last Exam, a test of PhD-level scientific reasoning.
Thinking Machines is aware of these limitations. While Inkling is not the most powerful model available, either open or closed, it is the most capable open-weights model developed by a Western lab. Developers who are constrained by legal, security, or compliance issues from utilizing models created in Beijing now have a legitimate alternative to self-hosting Chinese models.
This gives developers a model that, although it may fall short compared to the best Chinese alternatives, aligns more closely with their values, expectations, and ideologies. Future fine-tuning efforts could enhance this model's performance for specific tasks, making it competitive in benchmarks against Asian models.
Inkling also scored 78.0% on FORTRESS Adversarial, which evaluates how effectively a model resists harmful prompts while minimizing the blocking of legitimate content, marking the highest score among all open-weights models in this comparison.
In addition to Inkling, Thinking Machines previewed Inkling-Small, which features 276 billion total parameters and 12 billion active. This smaller model is already matching the larger version on most reasoning benchmarks, with its weights expected to be released after testing, although no timeline has been provided.