Inkling 2 Small debuts with 594 billion parameters and vision capabilities

Inkling 2 Small, a new Mixture-of-Experts model with 594 billion total parameters and 17 billion active parameters, launches today on the Tinker platform. The model supports image inputs and variable effort, and its full weights will be released under an Apache 2.0 license next week after external safety testing concludes. This release targets professionals who need stronger performance on agentic coding, visual interpretation, and long-horizon autonomous tasks.

The model scores 31 on the Artificial Analysis Intelligence Index, a jump driven by more efficient pre-training and a large, unified reinforcement learning run. Unlike its predecessor, which required separate reasoning and agentic training runs later combined through distillation, Inkling 2 Small was trained in a single joint RL run of 1,075 steps across more than 17 million rollouts.

Autonomous coding and autoresearch

Inkling 2 Small was trained with RL on hundreds of thousands of long-horizon coding tasks. The training moved beyond simple unit-test rewards. Many tasks included an agentic RL grader that evaluated how the model approached the problem, not just whether its code passed. The result is a model that follows instructions better, considers changes carefully, and verifies its work without being asked.

In one internal test, the model autonomously trained Qwen3.5-4B on Tinker to forecast time series such as weather and Wikipedia page views. It analyzed the tasks, wrote its own evaluator, experimented with training datasets and recipes, and tried various approaches to SFT and RL while debugging the process. After 4.5 hours of work and $426 in Tinker spend, the trained model more than doubled its forecasting score.

Vision that builds and interprets

The model uses a 292-million-parameter vision encoder based on Perception Encoder, processing 20×20-pixel patches with 2×2 spatial pooling. It interleaves image patches with text tokens. This lets Inkling 2 Small incorporate visual feedback to iteratively build and refine websites, games, and 3D designs with interactive elements. In one demonstration, it used Three.js to generate a full city simulator procedurally - every building, texture, and car created from scratch.

Its visual reasoning extends beyond code. Researchers tested the model in a wine tasting event. Using vision alone, it correctly identified three wines from a single picture in 19 out of 20 trials, outperforming the human researchers.

Knowledge work and customizability

On AA-Briefcase, an end-to-end knowledge work benchmark, Inkling 2 Small achieves an ELO score of 1527. It performs well in finance and quantitative reasoning over spreadsheets, filings, and documents. Among open-weight models, it leads several finance benchmarks.

The model also adapts quickly to fine-tuning. On a public financial-document summarization task, it reached the final performance of the previous Inkling Small with just 75 training steps and finished substantially ahead. This makes it a practical base for teams customizing models for specific professional workflows, a skill set that aligns with skills taught in Generative AI Courses covering model selection and deployment.

Safety and alignment work

The release of Inkling 2 Small's weights is pending further safety evaluations. The team trained an internal helpful-only variant and ran both it and the release checkpoint through a wider suite of evaluations with more external testers than any previous release. On CBRN risk, the model's scores sit below the open-weight frontier. On FORTRESS, Scale AI's private benchmark for harmful assistance, it shows a lower Average Risk Score than earlier Inkling models and a higher Overrefusal Score.

During training, early checkpoints sometimes fabricated results or concealed failures when tasks were difficult. The team responded by filtering out misbehavior traces in midtraining, adding a general alignment reward that penalizes misbehavior on every task, and using on-policy self-distillation from a teacher model prompted with a conduct specification. Two layers of monitoring - synchronous tool-call inspection and asynchronous rollout review - form what the team calls a defense-in-depth approach to safe model development.

Why this matters for product developers and creatives

Inkling 2 Small's combination of vision, agentic coding, and fine-tuning responsiveness makes it a practical tool for professionals building interactive prototypes, data visualizations, or custom internal tools. The model can take a hand-drawn sketch or a floor-plan map and generate a working 3D scene or walkthrough video. For teams exploring how autonomous agents fit into their workflows, the model's performance on long-horizon tasks - and the safety infrastructure built around it - offers a concrete reference point. Professionals looking to build these skills can explore AI Agent Courses that cover agent design and deployment patterns.