Snowflake Adds Inkling 2 Small to Cortex Training
Snowflake has made Inkling 2 Small available in Cortex Training public preview, giving teams a way to benchmark, post-train, evaluate and serve the open-weight model on their own enterprise workloads. The model comes from Thinking Machines Lab and can be used directly where customers’ data lives, under Snowflake’s existing governance controls.
- Inkling 2 Small is now available in Snowflake Cortex Training public preview.
- The model has 594 billion total parameters and 17 billion active parameters.
- It supports image inputs, variable reasoning effort, post-training and inference.
- Its weights are released under the Apache 2.0 license.
- Snowflake says teams can use it on governed enterprise data and context.
An open-weight model for agentic work
Inkling 2 Small is the second model generation from Thinking Machines Lab. Snowflake describes it as a mixture-of-experts, or MoE, model with 594 billion total parameters and 17 billion active parameters. It supports image inputs and variable reasoning effort.
Thinking Machines Lab trained the model with reinforcement learning on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses. According to the company, Inkling 2 Small follows instructions more closely, considers changes carefully and verifies its own work without being asked. Snowflake says these capabilities are intended to make the model easier to use inside real agent loops.
The model’s reported coding results include 86.3% on SWE-Bench Pro v2 (Hard) and 72.1% on DeepSWE v1.1. These figures are reported in Snowflake’s announcement and are presented as part of Inkling 2 Small’s coding performance.
Image inputs and Apache 2.0 licensing
Inkling 2 Small is multimodal, allowing it to process image inputs. Snowflake gives the example of using visual feedback to iteratively build and refine interfaces and designs. This extends the model’s use beyond text-only coding and reasoning tasks.
Its MoE design activates 17 billion parameters per token. Snowflake says this architecture delivers high capability at a lower inference cost than a dense model of comparable size. The model’s weights were released under the Apache 2.0 license, which Snowflake says allows enterprises to keep full ownership of models they customize.
What teams can do in Cortex Training
Cortex Training is Snowflake’s serverless offering for customizing and deploying open-weight large language models. With Inkling 2 Small, teams can use the service for post-training and inference. Snowflake says the model is available for both today, alongside benchmarking, evaluation and serving on enterprise workloads.
Cortex Training supports several customization methods, including supervised fine-tuning, reinforcement learning, direct preference optimization and distillation. It also integrates with major post-training frameworks and enables custom loss functions and architectures.
Snowflake says Inkling 2 Small joins a catalog that already includes open-weight model families such as Qwen, GLM, Nemotron, Gemma and DeepSeek. The service is designed to let enterprises customize and deploy these models on governed enterprise data and context, directly within Snowflake.
Why the availability matters
The release gives enterprises a way to test and adapt Inkling 2 Small without moving their data outside the environment where it is governed. Snowflake positions this as a response to growing enterprise interest in customizing open-weight models with proprietary data and context.
For users, the public preview combines access to Inkling 2 Small with Cortex Training’s post-training and deployment tools. Snowflake says enterprises can use the service to pursue higher accuracy and lower costs by customizing open-weight models for their own workloads. The announcement does not provide specific pricing or a general availability date.
Snowflake: Inkling 2 Small Now Available in Snowflake Cortex Training


