Inkling 2 Small arrives on Tinker, with downloadable weights still planned

What is available at launch

Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026. Its announcement offers Tinker Playground access plus fine-tuning and sampling through Tinker, initially with closed weights.

The company says it intends to release the full weights under Apache 2.0 the following week, after external safety testing and red-teaming. Downloadable checkpoints are tied to that weight release. This is a stated plan, not a completed download release in the launch announcement.

For a team starting an evaluation, that distinction affects the decision timeline. Hosted trials can answer questions about task quality now; a plan that depends on moving a checkpoint elsewhere needs a separate availability check.

Public record: Thinking Machines Lab launch announcement — November 5, 2026

Small describes the model tier, not a tiny download

The model card lists a mixture-of-experts architecture with 594 billion total parameters and 17 billion active. It accepts text and images and produces text, with a context window of up to one million tokens. The launch post explicitly says this model does not support audio.

The card lists Tinker Terms of Use as the current license. Its documented limitations include hallucinations, imperfect instruction following and degraded performance in long conversations.

Active and total parameter counts answer different questions. The active count describes selective computation; it should not be treated as the complete size of the model. Likewise, a large advertised context window does not establish accuracy on your longest document or conversation.

Public record: Inkling 2 Small model card — checked November 8, 2026 · Thinking Machines Lab launch announcement — November 5, 2026

Effort is a setting to test with the task

Tinker’s Inkling guide names thinkingmachines/Inkling-2-Small for Cookbook workflows. It recommends comparing learning rates and thinking effort, and provides a script for comparing standard effort values.

Our suggested first test is a small set of tasks with clear acceptance criteria. Include a screenshot that requires an accurate explanation, a code change with a checkable outcome, and a longer conversation that must preserve a specific constraint.

Record the accepted result, elapsed time, usage and corrections at each effort setting. Keep the same inputs and criteria across runs. That gives you a useful comparison before introducing a fine-tuned variant, where changes in both the model and the evaluation can otherwise become hard to separate.

Public record: Tinker: Using Inkling — checked November 8, 2026

Read the release as an evaluation opportunity

The practical next step is to decide what improvement you need: fewer corrections, better visual interpretation or more consistent behaviour in a particular workflow. A trial should test that requirement directly.

For self-hosting plans, return to the actual weight files and their license when they are released. For hosted use, assess the service and its current terms. Keeping those decisions distinct helps a team turn a model announcement into an implementation choice it can verify.