Inkling 2 Small pushes Thinking Machines up the rankings

Thinking Machines launches Inkling 2 Small, a 594-billion-parameter model accessible via Tinker. Its gains in coding and agent tasks come with a higher execution cost.

Thinking Machines has unveiled Inkling 2 Small, its second AI model, now available via its Tinker platform for inference and fine-tuning. This model with 594 billion parameters, including 17 billion active per query, stands out for its advances in coding, visual reasoning, and agentic tasks, but it also costs significantly more to use. The release is twofold: it shows that Mira Murati's company is making technical progress while reportedly, according to Business Insider, being close to a $5 billion fundraising round.

A model more open on paper than in practice

Inkling 2 Small is not yet an open model in the classical sense. Thinking Machines first offers it as a proprietary version on Tinker and Tinker Playground, before a planned release of the weights under an Apache 2.0 license "next week," after external security tests and red-teaming. The previous Inkling was launched with its weights available from the start. This change of approach clearly reflects the current strategy: move fast, then open the door after verification.

The model accepts text and images, with variable reasoning effort. Thinking Machines also claims it can use visual feedback to build and refine websites, games, and 3D designs. In this area, the positioning clearly targets practical uses, not lab demonstrations.

Des gains nets sur les tâches agentiques

Results reported by Artificial Analysis place Inkling 2 Small at 31 points on the Artificial Analysis Intelligence Index v5.0.1, 18 points higher than its predecessor. On agentic tasks, the progress is marked: AutomationBench-AA rises from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and AA-Briefcase reaches 1,527 Elo compared to 912 for Inkling Small. These benchmarks respectively measure workflow automation, software engineering, and long office tasks such as spreadsheets, presentations, and notes.

The model also proves more robust on certain creative and scientific tasks. It climbs to 56% on CritPt v2, up from 34% previously, and improves its results on AA-3D and Code2Video-AA. However, the picture is not uniform: on Terminal-Bench 4.0, Inkling 2 Small remains behind Claude Sonnet 5.5. In other words, the model advances but does not win every round at the benchmark table.

More talkative, therefore more costly to run

The performance increase comes at a real cost. Inkling 2 Small generates about 150,000 output tokens per task on the Intelligence Index, compared to 37,000 for Inkling Small. According to Artificial Analysis, the average cost rises to $1.18 per task, nearly six times more than its predecessor. The pricing displayed by Thinking Machines is $0.30 per million input tokens and $1 per million output tokens, with $0.03 for hidden input tokens.

This verbosity also has a side effect on hallucinations. The rate of fabricated responses drops from 61% to 27%, but Artificial Analysis notes that the improvement mainly comes from a model that attempts to answer less often: the attempt rate falls from 77% to 53%, while accuracy only rises from 32% to 34%. The result is therefore cleaner but not miraculous.

A valuation rising as fast as the scores

The new model version arrives as Thinking Machines is reportedly negotiating a nearly $5 billion fundraising round, according to Business Insider, led by Accel at a pre-money valuation of $40 billion. If the deal closes on these terms, the company would reach about $45 billion post-investment. This is a spectacular leap compared to the July 2025 round, when the startup raised $2 billion on a $12 billion valuation. The financing is not closed at this stage [to be verified].

This trajectory relies partly on a team from OpenAI and on a announced partnership with Nvidia, which is to provide at least one gigawatt of Vera Rubin systems starting in 2027. In short, Thinking Machines is trying to raise computing power, technical results, and company value simultaneously. The market is watching all this very closely.

Key points

  • Inkling 2 Small reaches 31 points on Artificial Analysis's AI Index.
  • AutomationBench-AA progresses from 5% to 46%.
  • The average cost rises to $1.18 per task.
  • Business Insider mentions a $5 billion fundraising round.
  • Thinking Machines aims for an Apache 2.0 release "next week."
  • Mira Murati founded Thinking Machines after leaving OpenAI.

By the numbers

  • 594 billion — total model parameters, with 17 billion active.
  • 8.4 trillion — tokens processed during initial pretraining.
  • 2,048 GPUs — used for the reinforcement learning phase over nine days.
  • $5 billion — amount of the fundraising mentioned by Business Insider.
  • $45 billion — potential post-money valuation if the deal succeeds.

To read