Thinking Machines Lab debuts Inkling 2 Small, a MoE model with 594B total and 17B active parameters, and plans to release weights next week after safety testing

The weights for Inkling 2 Small will be publicly available next week, enabling broader experimentation and fine-tuning.

Thinking Machines Lab has launched Inkling 2 Small, a new Mixture-of-Experts (MoE) model featuring 594 billion total parameters and 17 billion active parameters. This model demonstrates enhanced capabilities in agentic coding, autoresearch, and vision interpretation, outperforming previous models on the Artificial Analysis Intelligence Index with a score of 31. Inkling 2 Small is currently accessible via the Tinker platform with closed weights. The lab plans to release the full model weights under an Apache 2.0 license next week, contingent on the successful completion of external safety testing and red-teaming. The model's improvements stem from optimizations in pre-training and a comprehensive reinforcement learning run, enabling it to handle complex, long-horizon autonomous tasks and generate intricate outputs such as a procedural city simulator and a Python-based Linux PC emulator.

Inkling 2 Small represents a significant step in agentic AI, showcasing advanced capabilities in complex task execution, including code generation, autoresearch, and visual understanding. Its architecture, employing a Mixture-of-Experts approach with a substantial number of total parameters, allows for nuanced and efficient processing, particularly for long-horizon tasks. The model's performance on benchmarks like the Artificial Analysis Intelligence Index and specialized coding tasks suggests a leap in practical applicability for AI agents.

The release strategy, beginning with limited access on Tinker and followed by an open-weight release, balances rapid deployment with safety considerations. This approach allows for community-driven innovation and application development while mitigating potential risks through rigorous testing. The availability of downloadable checkpoints will further empower developers to fine-tune and integrate Inkling 2 Small into their own projects, potentially accelerating advancements in AI-powered tools and services.