How to Fine-Tune Inkling AI
Customization is Inkling's whole pitch. Here is the practical path from objective to deployed checkpoint on Tinker — including what the famous self-fine-tuning demo actually did.
Why Fine-Tuning Is the Point of Inkling AI
Most model releases pitch benchmark wins. Inkling AI's pitch is different: a broad, balanced multimodal base that you shape into your model. Thinking Machines built its business around this — the Tinker platform offers managed fine-tuning with Inkling supported from day one, and the launch's signature demo had Inkling fine-tune itself into a lipogram model that never writes the letter "e".
That demo is worth studying not as a stunt but as a template: it shows the exact loop — objective, synthetic data, rubric, training job, evaluation, weight swap — that any real Inkling AI fine-tune follows. Below is that loop as a practical checklist.
The Five-Step Inkling Fine-Tuning Loop
Pick your objective
Define the behavior you want that prompting alone can't reliably achieve — a domain-specific voice, a strict output format, or specialized knowledge. Thinking Machines' launch demo used a playful one: a lipogram model that never writes the letter 'e'.
Prepare data and an eval
You need training examples and a scoring function. In the official demo, Inkling generated its own synthetic data and a rubric (score 0 if 'e' appears, 10 otherwise). For real use cases, a few hundred high-quality examples beat thousands of noisy ones.
Launch the job on Tinker
Tinker is Thinking Machines' managed fine-tuning platform with day-one Inkling support. Jobs are defined in Python with the tinker_cookbook library — the demo ran 32 batches × 3 epochs (96 steps) and finished in about 27 minutes.
Evaluate against the base
Always compare the fine-tuned checkpoint against base Inkling on your eval before switching. Tinker logs metrics per step so you can catch regressions early.
Deploy the new checkpoint
Checkpoints get a tinker:// URI you can load into your serving setup or agent harness. The demo's self-update hot-swapped weights inside OpenCode without restarting the session.
What the Self-Finetune Demo Actually Proved
Numbers from the official run: Inkling wrote ~230 lines across three files (objective, training config, self-update script), trained for 96 steps, finished in about 27 minutes, and passed its own eval before hot-swapping weights inside the OpenCode harness. Context used: 43K tokens of a 1M window.
Two practical takeaways. First, small targeted fine-tunes are fast and cheap — this is an hours-not-days workflow. Second, the eval matters more than the training: the demo's rubric (score 0 if "e" appears, 10 otherwise) is trivially checkable, which is exactly what made the loop trustworthy. Design your scoring function before your dataset.