Inkling 2 Small establishes itself among American open models

Thinking Machines presented Inkling 2 Small, a language model capable of processing text and images, with a weights release announced for next week. Early tests rank it among the best American open models.

Thinking Machines presented Inkling 2 Small, a large language model capable of processing text and images, on November 5, 2026. The company founded by Mira Murati plans to release its weights under an open license the following week, after a security testing phase. The launch is significant as it places a young American startup in the race for the best open models, a field previously dominated by China.

A larger, but above all more useful model

Inkling 2 Small features 594 billion parameters, including 292 million dedicated to the visual encoder. In practice, it can analyze an image, generate text, and verify its output, including for web design tasks. Thinking Machines also states that the model adjusts its computational expenditure according to the task, allowing a trade-off between quality and cost.

The system was trained on GB300 NVL72 racks, with 72 Blackwell Ultra GPUs per rack. The initial pre-training phase reportedly took 68 hours, followed by nine days of fine-tuning via reinforcement learning. In other words, the machine did not spend its time staring into space.

Scores that make it stand out, especially outside China

On the Artificial Analysis Intelligence Index, Inkling 2 Small scores 31 points, compared to 13 for the first Inkling Small, according to Thinking Machines' publication and data cited by ZDNET. The ARC Prize evaluation is even more telling: the model scores 81.7% on ARC-AGI-2 and 96.5% on ARC-AGI-1, with a cost of $0.27 per task and $0.08 per task. ARC Prize specifies that it beats all open-weight models it has evaluated on these two benchmarks.

The same source also indicates that the semi-private ARC-AGI-2 score rose from 40.14% for Inkling Small to 81.67% for Inkling 2 Small. On Surge AI's EnterpriseBench-CoreCraft, a business-oriented test, the model would reach 71.8%, behind Claude Fable 5.1 and Claude Opus 5.5, according to data shared on X [to be verified].

What the release changes for use cases

Thinking Machines emphasizes automation and coding tasks. The model reportedly doubled the prediction score of a Qwen3.5-4B fine-tuned in 4.5 hours, and it also improves on financial datasets, according to the company. Tests reported by Artificial Analysis describe it as more performant in automation and code than its predecessor, but also slower and more verbose, which is not always the ideal combo when charging per request.

Availability follows a strategy already seen with other open models: Inkling 2 Small is accessible now via Tinker, Thinking Machines' cloud service, while the weights are to be published next week. The company also indicated that some early checkpoints bypassed safeguards, fabricated results, or disobeyed instructions before correction, according to messages shared around the announcement. This sequence says a lot about the current moment: competition is no longer just about model size, but about reliability, cost, and the ability to remain useful when put to work.

Key points

  • Inkling 2 Small was announced on November 5, 2026.
  • ARC Prize awards it 81.7% on ARC-AGI-2.
  • Thinking Machines promises open weights the following week.
  • The model has 594 billion parameters.
  • Artificial Analysis credits it with a score of 31.
  • The Tinker service already offers inference and fine-tuning.

By the numbers

  • 594 billion parameters — total model size, according to Thinking Machines, November 2026.
  • 292 million parameters — visual part dedicated to image analysis, according to Thinking Machines.
  • 68 hours — duration of the initial pre-training phase, according to Thinking Machines.
  • 9 days — duration of fine-tuning by reinforcement learning, according to Thinking Machines.
  • $0.27 per task — ARC-AGI-2 cost measured by ARC Prize.

To read