Inkling 2 Small in Cortex: Test the Whole Fine-Tuning Job, Not the Parameter Count

A model described as “small” can still contain hundreds of billions of parameters. The label makes sense only relative to its architecture and active computation. Snowflake’s November 12 addition of Thinking Machines Lab’s Inkling 2 Small to Cortex Training is a useful prompt for platform teams to ask a more practical question: what does it cost to adapt and serve the model on our workload, under our data controls?

What is available

Snowflake’s announcement says Inkling 2 Small is available in Cortex Training public preview for benchmarking, post-training, evaluation and serving. Thinking Machines Lab describes the model as a mixture of experts with 594 billion total parameters and 17 billion active parameters. It accepts image input and supports variable reasoning effort. Snowflake cites coding results from the model’s launch, but those are vendor-reported scores, not an independent comparison on a buyer’s environment. ^{[1] [2] [3]}

The difference between total and active parameters is relevant because a mixture-of-experts model routes a token through a subset of its components. It does not mean that the rest of the system disappears from memory or deployment planning. Our explainer on active parameters and local model trade-offs covers why that distinction matters for deployment choices, while our open-weight model release analysis looks at what an open licence does and does not settle. The weights, expert placement, routing behavior, hardware and serving stack all affect what a real deployment requires. A single active-parameter number cannot be used as a hosting-cost estimate.

Snowflake positions Cortex Training as a serverless service for customizing and deploying open-weight models near governed enterprise data. That can reduce some integration work, especially when data already lives in the platform. The exact data path, supported regions, preview terms and configuration limits should be confirmed in current Snowflake documentation before teams move sensitive datasets or make a production commitment. ^{[1] [4]}

Measure the whole job

A meaningful evaluation starts with a representative workload. Choose tasks that reflect the reason for customization: classification, structured extraction, code assistance, image understanding or a domain-specific agent step. Build a held-out set and compare the base model with the tuned model. Track error types, not only a single average score. If the model will call tools, include the full interaction and failure recovery in the test.

Measure the lifecycle, not just one inference. Include preparation of training examples, post-training compute, evaluation, deployment, cold starts, throughput and ongoing serving. Test realistic concurrency and context lengths. A model may look inexpensive on a small prompt and become costly when it must reason longer, return images or serve many simultaneous users.

Also include the cost of data movement and governance. Keeping information close to its source can simplify controls, but teams still need to inspect who can initiate training, how outputs are logged, whether training artifacts inherit access restrictions and how model versions are promoted. A managed platform can provide useful controls; it does not remove the need to assign ownership and review them.

Treat headline benchmarks as a starting point

The launch materials report scores on coding benchmarks and describe training choices intended to improve long-horizon agentic work. Those figures can help identify which tasks a vendor believes the model handles well. To compare models fairly, teams need the benchmark version, prompt and harness conditions, contamination safeguards, scoring rules and whether the same amount of inference effort was allowed. Public scores rarely capture the operational constraints of one company’s workflow. ^{[2] [3]}

A useful decision memo should show at least three things: quality on the team’s own held-out examples, cost per successfully completed task, and the failure modes requiring human review. “Per token” is not the right unit when one model call succeeds in one attempt and another requires retries, longer reasoning or repair. Calculate the cost of a usable result.

Preview means evaluate before depending

Public preview is an invitation to learn, not a service-level promise. Confirm current availability and constraints, preserve a rollback path to the existing model, and avoid placing a new model on a critical production route until it passes the same data, quality and reliability gates as its alternatives.

Inkling 2 Small’s addition gives Snowflake customers another open-weight model to test within Cortex Training. The practical value will come from fit: whether the model’s capabilities justify the full training and serving footprint for a defined job. Parameter counts and vendor benchmarks can narrow the search. Only a workload-level evaluation can make the buying decision.