Thinking Machines Launches Competitive Open-Weight AI Model
Inkling 2 Small challenges Chinese dominance in the open-weight artificial intelligence sector.
1New AI Market Entrant
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has unveiled Inkling 2 Small. This new model is designed to compete directly with leading Chinese open-weight models that have recently dominated the sector. The company plans to release the model weights publicly next week following rigorous internal safety and security evaluations.
2Technical Model Specifications
The model features 594 billion parameters and supports image inputs while producing text outputs. It has demonstrated significant improvements in web and game design capabilities. According to the company, the architecture is specifically optimized to perform better at coding tasks and function more effectively within complex real-world agent loops for developers.
3Benchmark Performance Metrics
Inkling 2 Small achieved a score of 31 on the Artificial Analysis Intelligence Index, which tracks progress toward artificial general intelligence. This performance ties it with Gemini 3.8 Flash. While it remains behind industry leaders like Claude Opus 5.5, it currently stands as the highest-scoring open-weight model developed outside of China.
4Staggered Release Strategy
Thinking Machines is adopting a staggered release process, similar to the approach used by French lab Mistral for its Le Chonk model. This method involves spending several weeks conducting cybersecurity audits and safety reviews before making the weights available. This practice is becoming standard to address heightened global concerns regarding AI safety.
5Performance and Cost Analysis
Artificial Analysis reviewers noted that while the model is highly intelligent, it is notably slow per task and quite verbose. Furthermore, at $0.30 per one million input tokens, it is priced above the $0.24 median for its class. These factors present a trade-off for enterprise customers evaluating their AI infrastructure costs.
6Significant Capability Gains
The model shows substantial progress compared to its predecessor, with a 41% increase in automation capabilities and a 20% improvement in coding performance. These gains suggest that Thinking Machines is rapidly closing the gap with more established proprietary labs, despite the challenges inherent in balancing speed, cost, and overall model intelligence.
7Accuracy and Hallucination Rates
Testing on the AA-Omniscience v1.1 benchmark revealed meaningful improvements in reliability. The hallucination rate dropped significantly from 61% to 27%, while accuracy rose from 32% to 34%. These metrics indicate that the new model is more dependable for complex tasks, even if the overall attempt rate has decreased to 53%.
8Shifting Competitive Landscape
The emergence of Western-made open models like Inkling 2 Small forces investors and enterprise clients to reconsider their reliance on expensive, closed-source contracts. As Western labs increase their output of competitive open-weight alternatives, the market dynamic is shifting away from the total dominance previously held by Chinese models in this specific category.
Original article: ZDNET — Thinking Machines Launches Competitive Open-Weight AI Model