Inkling 2 Small from Thinking Machines Lab Is Now Available in Snowflake Cortex Training

Enterprises are increasingly turning to customizing open-weight models on their own data and context to have full control over the behavior and results. 

Today, we are excited to share that Inkling 2 Small, the latest open-weight model from Thinking Machines Lab, is now available in Snowflake Cortex Training (public preview). Teams can start benchmarking, post-training, evaluating and serving Inkling 2 Small on their own enterprise workloads today, directly where their data lives and governed by the same controls they already rely on.

What is Inkling 2 Small?

Inkling 2 Small is the second model generation from Thinking Machines Lab. It is a Mixture-of-experts (MoE) model with 594 billion total parameters and 17 billion active parameters, and it supports image inputs and variable reasoning effort. 

Inkling 2 Small is built for long-horizon agentic work. Thinking Machines Lab trained it with reinforcement learning (RL) on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses, and it reports that the model follows instructions more closely, considers changes carefully and verifies its own work without being asked. That makes it easier to use inside real agent loops, and it shows in its coding results, including 86.3% on SWE-Bench Pro v2 (Hard) and 72.1% on DeepSWE v1.1.

The model is also multimodal. Image inputs let it incorporate visual feedback, for example to iteratively build and refine interfaces and designs. And because only 17 billion of its parameters are active per token, the MoE architecture delivers high capability at a lower inference cost than a dense model of comparable size. With weights released under the Apache 2.0 license today, enterprises keep full ownership of the models they customize.

What can you do with Inkling 2 Small on Snowflake?

Cortex Training, Snowflake’s serverless offering for customizing and deploying open-weight LLMs, is designed to make open-weight model customization simple, flexible and cost efficient. We are thrilled to welcome the Inkling family of models to Cortex Training’s rich catalog of existing open-weight models, including families such as Qwen, GLM, Nemotron, Gemma and DeepSeek. Inkling 2 Small is available for both post-training and inference today.

Figure 1: Cortex Training supports a wide range of models and training methods — including supervised fine tuning, reinforcement learning,  direct preference optimization and distillation — while integrating with major post-training frameworks and enabling custom loss functions and architectures.

Get started

As open-weight models become increasingly competitive with frontier models, enterprises have an opportunity to customize AI for higher accuracy and lower costs. With Cortex Training, enterprises can flexibly and cost-effectively post-train and deploy open-weight models on their governed enterprise data and context. To learn more about the latest innovations in Cortex Training, check out the launch blog post, github repo, and web page.