Thinking Machines releases Inkling 2 Small: AA score 31, top US open-source model

Beating AI News Flash: Thinking Machines, founded by former OpenAI CTO Mira Murati, has released a new model, Inkling 2 Small. It achieved a composite score of 31 on Artificial Analysis, up from only 13 for the previous generation, making it the highest-scoring open-source model in the United States.

The model is named Small, but it actually has 594 billion total parameters, with 17 billion activated per inference. The biggest improvement is in Agent capabilities. The automated task pass rate rose from 5% to 46%, and the coding test score rose from 5% to 25%, though it is still slightly below DeepSeek V4.1 Flash's 28.5%. Thinking Machines mainly relies on reinforcement learning to improve the model's ability to handle complex tasks.

Artificial Analysis testing found that the model outputs about 150,000 tokens per task on average, 4 times that of the previous generation. The average cost rose from $0.20 to $1.18, and completing a task takes about 21 minutes.

Currently, Inkling 2 Small has opened inference and fine-tuning on Tinker. The full weights have not yet been released. The official plan is to complete external security testing next week and then open-source it under the Apache 2.0 license.