Inkling 2 Small AI Model: Features, Ranking and Capabilities

Thinking Machines Lab has introduced Inkling 2 Small, a multimodal AI model boasting 594 billion parameters that has quickly climbed to the top tiers of web design and coding benchmarks. Operating under the leadership of Mira Murati, the lab's new release showcases advanced capabilities in handling both text and visual inputs.

• Inkling 2 Small scored 1340 points on the DesignArena website generation leaderboard.
• The Mixture of Experts (MoE) architecture activates only relevant parameters per query.
• The model uses visual feedback to create and debug websites, games, and 3D projects.
• Model weights will be released under the Apache 2.0 license following safety audits.

Architecture and Core Capabilities

Inkling 2 Small relies on a Mixture of Experts (MoE) architecture, which helps manage its massive scale by activating only a fraction of its parameters for any given request. While still requiring substantial computational resources, this design optimizes efficiency compared to fully dense models and supports adjustable reasoning depth.

One of the standout features of the model is its use of visual feedback during programming. It can generate websites, games, and 3D projects, subsequently analyzing screenshots of the rendered output to autonomously spot errors and apply necessary code fixes.

Benchmark Performance and Training

In the November 5 DesignArena website creation benchmark, Inkling 2 Small secured the 3rd position with 1340 points. It trailed only Muse Wonder (1366 points) and Muse Spark 1.3 (1351 points), while outperforming several prominent systems like GPT-6 Astra, Kimi K3, and Claude Opus 5.5.

During its training phase, the model was exposed to prolonged software engineering tasks. This included handling real code repositories, optimizing graphic processing units (GPUs), and even participating in the autonomous training of other AI models.

Availability and Future Roadmap

Currently, Inkling 2 Small is accessible for execution and fine-tuning through the Tinker platform, though its core model weights remain private for now. Thinking Machines Lab plans to release the weights under an open Apache 2.0 license next week.

The public release is scheduled immediately following the completion of external safety checks and robustness tests against prompt injection and jailbreak attempts. Until then, developers can access and evaluate the model securely through the lab's proprietary infrastructure.