Fine-Tune Inkling, the Open 975B Multimodal Model, on IP.ai

2026-07-16 · 6 min read

On July 15, 2026, Thinking Machines Lab released Inkling, a 975B-parameter (41B active) open-weight, multimodal Mixture-of-Experts model that reasons over text, images, and audio, with controllable "thinking effort." It's Apache-2.0 licensed, and the full weights are on Hugging Face. IP.ai now integrates Tinker, Thinking Machines' training API, so you can fine-tune Inkling (and other very large open models) on your own data.

Why this matters for the rest of us

For the trades, for researchers, for small teams, the pitch has always been the same: own your data, own your model, keep costs low, and never be forced to become an ML engineer. Inkling widens what "own your model" can mean:

• Multimodal, open, and yours. Inkling takes text, images, and audio in. Fine-tuned on your material, the adapter is Apache-2.0 and exportable on request, with no shared-model lock-in.
• Efficient by design. Its controllable thinking effort lets you trade a little quality for big savings in latency and cost. That's the difference between a demo and something you run thousands of times a day.
• Sizes past what fits on one GPU. Through Tinker, the training happens on Thinking Machines' distributed infrastructure, so model sizes from 1B to 1T+ are on the table, not just what squeezes onto a single card.

How IP.ai fits in

You still don't touch any of the machinery. In the model wizard you:

1. Upload your data (we clean and prepare it).
2. Pick Inkling and the Tinker training provider.
3. Describe your goal and hit train.

Under the hood we talk to the Tinker API on your behalf. Tinker's own docs describe fine-tuning as a clean loop (forward_backward, optim_step, save_weights) with the heavy distributed-GPU lifting handled remotely. IP.ai wraps that so you never see it. And because Inkling is open-weight, you can export the trained checkpoint and run it anywhere.

Choose your ground

Inkling and Tinker join the training options you already have on IP.ai:

Same data, your call. And you can still switch on delete-my-data-after-training with any of them.

It's live

The Tinker integration now runs end to end: a real LoRA fine-tune on Thinking Machines' GPUs, a servable checkpoint, and inference through your OpenAI-compatible endpoint, all from the model wizard, no code. We proved it on Llama 3.2 3B (fine-tuned and generating in minutes), and the same path serves the giants: Inkling, Kimi K2.6, Qwen3.5-397B, and gpt-oss.

Want to fine-tune Inkling on your data? Start your free trial.