Inkling 2 Small powers Thinking Machines’ agentic AI

Thinking Machines has unveiled Inkling 2 Small, a second-generation AI model designed for autonomous coding and visual reasoning. However, its performance gains come with a higher usage cost.

Thinking Machines has introduced Inkling 2 Small, a second-generation AI model aimed at autonomous coding, visual reasoning, and assisted work tasks, via its Tinker platform. The system features a total of 594 billion parameters, 17 billion active, and supports images. Its appeal is clear: it advances agentic uses further, but it also costs more to run than its predecessor.

A more ambitious model, but not free in computation

Inkling 2 Small is based on a mixture-of-experts architecture, a setup where only certain subnetworks are activated depending on the task to save resources. Thinking Machines also states that the model can adjust its reasoning effort and use visual feedback to build or improve websites, games, and 3D designs. In other words, the tool is not limited to responding: it can chain actions. This is precisely what attracts companies and drives up the bill.

The company founded by former OpenAI CTO Mira Murati specifies that it trained the model on 8.4 trillion tokens during an initial phase, then on 3.8 trillion more recent data. This first stage ran for 68 hours on Nvidia GB300 NVL72 systems, followed by a nine-day reinforcement phase involving 2,048 GPUs and over 17 million rollouts. The message is clear: Thinking Machines is heavily investing to shorten the path from lab to product.

Clear progress on agentic tasks

Independent tests by Artificial Analysis align with the company’s claims. Inkling 2 Small achieves a score of 31 on its Intelligence Index, 18 points higher than the previous model, with its best advances on agentic tasks. On AutomationBench-AA, which measures automated business workflows, the score rises from 5% to 46%. On FrontierCode 2.0, focused on software engineering, it increases from 5% to 25%.

The comparison with competitors remains more nuanced. Inkling 2 Small outperforms Claude Sonnet 5.5 on FrontierCode 2.0, but the model lags behind on Terminal-Bench 4.0. Above all, the gains come at a price: Artificial Analysis estimates the average cost at $1.18 per Intelligence Index task, compared to $0.20 for Inkling Small, with about four times more output tokens consumed. Performance improves, so does the ticket price.

A technical release amid a closely watched financial sequence

The launch of Inkling 2 Small comes as Thinking Machines is also at the center of another news story: according to Business Insider, the company is close to a $5 billion funding round led by Accel, with a pre-money valuation of $40 billion. If the deal is confirmed, the post-money valuation would reach about $45 billion. At this stage, the company has not commented and Accel has not responded.

This sequence matters for the AI market because it shows how a young company can turn a research promise into highly valued technological assets, even before revealing much about its revenue. Thinking Machines had already raised $2 billion in July 2025, at a $12 billion valuation, then announced in March a partnership with Nvidia covering at least one gigawatt of Vera Rubin systems starting in 2027. The Inkling 2 Small model adds a piece to this story, without lifting any veil on profitability. Not yet, at least.

Key points

  • 594 billion parameters in total, 17 billion active.
  • Artificial Analysis: score 31, +18 points.
  • AutomationBench-AA: 46% versus 5% before.
  • Targeted funding round: $5 billion.
  • Accel would lead the operation, according to Business Insider.

By the numbers

  • 8.4 trillion tokens — initial pretraining of Inkling 2 Small, according to Thinking Machines.
  • 3.8 trillion tokens — more recent data added afterward, according to Thinking Machines.
  • 2,048 GPUs — resources mobilized during the reinforcement phase.
  • 17 million rollouts — simulations performed on software engineering, documents, and tools.
  • $1.18 — average cost per Intelligence Index task, according to Artificial Analysis.

To read