Thinking Machines Lab has unveiled Inkling 2 Small, a model that third-party tester Artificial Analysis says could become the highest-scoring western open-weight AI release to date. The catch: its weights are not public yet. Thinking Machines, the startup led by former OpenAI CTO Mira Murati, said it will publish them next week.
What the model brings
Inkling 2 Small has 594 billion parameters. It accepts inputs but produces text-only outputs. The company highlighted its strengths in web and game design, and said it tested how the model approached coding problems in addition to the code it generated.
Thinking Machines described the result as a model that is stronger at coding and more practical inside real agent loops.
The model earned a score of 31 on the Artificial Analysis Intelligence Index, a composite benchmark that tracks progress across coding, science, and general categories. That score is above average for models in its class and matches Gemini 3.8 Flash. For scale, Gemini 4 Carbon and Claude Opus 5.5 sit at the top of the index with scores of 60 and 59, respectively.
Why a western open model matters
For several years, Chinese developers have dominated the open-weight space. Models from DeepSeek and Moonshot have regularly topped leaderboards and matched far more expensive closed systems from OpenAI and Anthropic. Since DeepSeek’s R1 performed comparably to OpenAI’s o1 in January 2025, Chinese open models have given investors and enterprise buyers reason to reconsider costly contracts with US labs.
More competitive open releases from western labs could alter that calculation. Artificial Analysis noted that if Thinking Machines follows through on its weight release, Inkling 2 Small would rank as the highest-scoring open-weight model developed outside China based on current standings.
The staggered release is becoming standard. Companies now spend weeks testing a to-be-open model with cybersecurity teams and safety reviewers before publishing weights, rather than leading with them immediately. French lab Mistral followed the same process with its Le Chonk model, which now keeps pace with OpenAI’s GPT-6.1 Terra in coding — a rare feat for a western open model.
The trade-offs behind the score
Artificial Analysis did not describe Inkling 2 Small as a complete package. The organization called it among the leading models in intelligence and reasonably priced relative to peers of similar cost, but also notably slow per task and very verbose.
At $0.30 per one million input tokens, the tester labeled the pricing somewhat expensive, noting that the median cost for its class is $0.24.
Even so, the gains over its predecessor stood out. Inkling 2 Small improved 41% in automation capabilities and 20% in coding abilities compared with Inkling Small.
Accuracy also moved. On AA-Omniscience v1.1, the hallucination rate dropped from 61% in Inkling Small to 27%, while accuracy rose from 32% to 34% and the attempt rate fell from 77% to 53%.
Most models excel in specific areas or on particular benchmarks, which makes it difficult to crown any single system the best coding model overall. Inkling 2 Small’s value will depend on how it performs in real deployments once its weights are available — and on whether it can help shift the open-weight landscape away from a single region’s dominance.
