Inkling 2 Small

Agentic coding and autoresearch benchmarks are now measurably better, influencing which models are chosen for complex development tasks.

Thinking Machines Lab has released Inkling 2 Small, their second-generation Mixture-of-Experts model. This new model boasts 594 billion total parameters with 17 billion active, and it supports image inputs and variable effort. Inkling 2 Small demonstrates significant performance gains, scoring 31 on the Artificial Analysis Intelligence Index, an improvement attributed to enhancements in agentic coding, vision, knowledge work, and reasoning. These advancements stem from more efficient pre-training and a large, unified reinforcement learning run. Initially, Inkling 2 Small will be available via Tinker with closed weights, with plans to release the full weights under an Apache 2.0 license next week after external safety testing. The model is available for testing on the Tinker Playground and for fine-tuning.

Inkling 2 Small showcases notable improvements in agentic coding and autoresearch, outperforming several other models in benchmarks like DeepSWE and SWEBench Pro. Its ability to perform long-horizon autonomous work, including complex tasks like building a city simulator from scratch or creating a PC emulator, highlights a leap in practical AI application. The model's training, which incorporated agentic RL grading and real-world coding task assessments, suggests a focus on creating models that are not only powerful but also more user-friendly within agent loops, exhibiting better instruction following and self-verification.

The release strategy, offering closed weights initially followed by an Apache 2.0 release after safety testing, aligns with a cautious approach to broader AI model accessibility. This phased rollout allows for user interaction and fine-tuning on the Tinker platform while managing safety concerns. The model's performance in specialized benchmarks like kernel optimization and autonomous time-series forecasting indicates its potential to tackle advanced scientific and engineering problems, suggesting an evolving landscape where AI agents can autonomously conduct complex research and development.