Inkling 2 Small is live on Tinker. What does “17B active” actually buy?

Thinking Machines Lab introduced Inkling 2 Small on November 5, making inference and fine-tuning available through Tinker while initially keeping its weights closed. The company intends to release the full weights under Apache 2.0 the following week, after external safety testing and red-teaming. That is a conditional plan, not an already completed download release.

For developers, the useful questions are what can be run now, how much work each token requires, and which limits apply to the chosen service. The word “Small” cannot answer all three.

A sparse model still has a large total size

The model card lists 594 billion total parameters and 17 billion active. Its 48-layer mixture-of-experts decoder routes each token to eight of 384 experts, combining local and global attention. It accepts text and images and produces text, with a stated model context capacity of up to one million tokens.

An organ can have many pipes while sounding only a selection at once. Sparse routing similarly selects part of a larger system for a token; it does not turn the full collection of weights into a 17-billion-parameter model. The active count alone is insufficient to specify a self-hosting machine or its memory requirements.

Tinker's pricing definitions distinguish total model size from parameters used per token. The service's rates and supported configurations are more actionable for a hosted trial than an assumed hardware requirement inferred from the model name.

The Tinker overview separates the developer's training loop from the distributed GPU computation handled by the service. It currently implements low-rank adaptation, or LoRA, rather than full fine-tuning. That distinction matters: customizing a model through an API is not the same operation as downloading and updating every original weight yourself.

Editorial visualization

One million tokens is not every endpoint's limit

On November 7, Tinker's service documentation listed these configurations. The training catalogue has 64K and 256K context options; the separate serverless inference beta lists 256K. A model's advertised capacity does not override the configuration selected in an API.

The same page lists cached prefill at $0.03 per million tokens and checkpoint storage at $0.10 per GB per month. It asks prospective production users of serverless inference to join a waitlist. These are published unit rates and access conditions, not a complete budget for an agent that may retry, use tools or generate lengthy reasoning.

The staged approach predates this launch. In its July 31 release framework, Thinking Machines described hosted inference and hosted fine-tuning as ways to widen access while retaining monitoring and revocation controls. It explicitly said stages do not automatically culminate in open weights. That older policy explains the distinction between access and ownership; it is not a fresh release confirmation.

Read the harness beside the score

The model card's Terminal Bench v2.1 results show 85.1% for Inkling 2 Small with the Pi harness and 81.3% with Terminus. These are vendor-reported results for differently configured agent systems, not two interchangeable measurements of model quality. The harness is the surrounding software that supplies tools and manages the agent's work.

The launch evaluation notes specify effort 0.99, temperature 1.0 and a maximum trajectory of one million tokens, increased from the 256K limit used for Inkling Small. A trajectory budget describes the work allowed across a run; it should not be mistaken for the context retained in every individual request. Comparing generations therefore requires checking the evaluation setup as well as the score.

Editorial visualization

The model card also warns of hallucinations, imperfect instruction following and degraded performance in long conversations. UpThat has not independently run these evaluations. The launch documents establish what the lab reports and offers, rather than proving reliability on a reader's workload.

A useful trial would keep the task, tools and spending limit fixed, then record correct completions, retries and failures. For customization, reserve examples the model never trains on and check performance against the unmodified starting point. The decision is whether adaptation improves that held-out work enough to justify its cost—not whether the model name sounds compact.