Snowflake adds Inkling 2 Small to Cortex Training preview

Snowflake has made Thinking Machines Lab’s Inkling 2 Small available in Cortex Training, where customers can benchmark, post-train, evaluate and serve the open-weight model in a managed environment.

The addition expands the model roster for Snowflake’s managed training preview. Snowflake’s model catalog lists the canonical model ID thinkingmachines/Inkling-2-Small-MXFP8. The model supports inference and LoRA training, which adapts a limited set of model parameters. Inference has a maximum context of 256,000 tokens on eight GPUs, while LoRA training has a maximum context of 256,000 tokens. The catalog does not list full-parameter training for the model.

Cortex Training is a serverless customization service. Snowflake says a client workspace handles orchestration and data preparation, while its managed platform holds model weights and optimizer state and performs GPU-intensive training and generation. Training data can come from Snowflake tables, internal stages or Amazon S3 buckets connected through external stages.

Snowflake’s documentation labels Cortex Training an open preview available in all commercial AWS regions and says customers must enable the service before use. The company warns that model availability, APIs and specifications may change before general availability. It also documents batch inference and REST deployment for evaluation and testing, including OpenAI-compatible inference endpoints backed by dedicated GPUs.