JOURNAL / 2026.11.09

Inkling 2 Small arrives on Tinker: adapting the model before receiving its weights

Thinking Machines launches its second generation with hosted fine-tuning and inference. The opportunity lies in adapting behavior; weight release remains conditional, and useful gains need testing outside training.

Thinking Machines introduced Inkling 2 Small on November 5. It reports improvements in agentic coding, vision, and adaptability over its first generation. Fine-tuning is already available on Tinker, but it plans to release weights under Apache 2.0 the following week, after external safety testing. This is a material development for people who want to change how an agent works, though the available service and the promised artifact deserve separate treatment.

The model card describes a mixture of experts with 594 billion total parameters and 17 billion active. It accepts text and images and produces text. “Small” is relative to the family: active parameters do not represent all the memory a self-hosted deployment would need. The card currently points to Tinker's terms.

There is also a difference from the original Inkling: that model supported audio and launched with weights available. I would therefore avoid reading “second generation” as a uniform improvement in every modality and form of access. For a voice application, losing that input may matter more than better coding performance.

The integration documentation identifies the new model and recommends exploring learning rate and reasoning effort together. This raises a more interesting decision than picking a leaderboard winner: how much behavior it can learn from our examples, and how much work it needs to do during each response.

Consider an agent that turns support tickets into proposed changes. We could teach it examples of when to request information, which tests to run, and how to deliver a modification for review. My expectation would be that fine-tuning could help maintain that procedure without repeating it in full in every instruction. But learning the shape of examples can also produce responses that look correct while skipping checks. The evaluation set should include cases where stopping or acknowledging missing evidence is the right decision.

The decisive comparison would be the unadapted model against the adapted model, with the same work budget. I would reserve tickets that played no part in training and check outcomes against criteria defined before fine-tuning. I would also test a product or policy change: an agent that learned yesterday's procedure well may apply it confidently after it has become invalid. This is an evaluation proposal, not an observed result for Inkling.

Pricing introduces another distinction. Tinker's table separates training, generation, and storage; serverless inference is listed as beta, with a production waitlist. A per-token rate alone cannot tell us what adapting and operating the agent will cost. My budget would include preparing examples, repeating experiments, and reviewing errors alongside service usage. Cheaper responses would not compensate for a model that constantly needed retraining.

Thinking Machines' openness policy presents hosted fine-tuning as an intermediate stage: it enables customization while the provider retains monitoring and the ability to revoke access. It also says these stages do not automatically lead to publishing weights. This policy helps explain the launch, but does not turn an intention to open into a contractual guarantee.

My reading is that Inkling 2 Small expands a valuable option: experimenting with specialized behavior before taking on model operations. For a small team, that may make it easier to discover whether fine-tuning is worthwhile. In exchange, the experiment depends on the provider's terms. I would proceed with repeatable tests and evaluation data retained by the team; I would postpone any decision requiring self-hosting until the weights, license, and actual requirements can be checked. The launch's value lies in what it lets us learn and improve today, without treating tomorrow's openness as already delivered.

Primary sources consulted on November 9, 2026. We have not run or fine-tuned Inkling 2 Small. Examples, expectations, and evaluation criteria are our own analysis.

← Back to journal