Inkling open-weights model is a 975B parameter multimodal MoE with 41B active parameters, Apache 2.0 licensing, and up to a 1M-token context window. That is the part engineers should care about. Not because it is the best model available, but because genuinely permissive open weights at this size are still rare.
This looks like a serious base model release. According to the source, it was pretrained on 45 trillion tokens across text, images, audio, and video, and Thinking Machines is positioning it very explicitly as a customization target rather than a frontier leaderboard play. In plain terms: they want you to fine-tune it, not just benchmark it.
That matters if you build internal AI systems, especially the kind of Microsoft Copilot and AI agents or custom workflows that need a model you can adapt to a narrow domain instead of renting whatever a closed vendor exposes this month.
What is actually interesting here?
The two strongest parts of this release are the license and the intent.
Apache 2.0 is the real story. A lot of so-called open AI models come with usage restrictions, revenue caps, or enough legal ambiguity that legal review becomes the actual blocker. This one appears much cleaner. If you care about portability, or just want the option to run your own stack later, that is a meaningful difference.
The second interesting part is that Thinking Machines is not pretending this is the strongest model on the market. That is refreshingly honest. They describe Inkling as a strong multimodal base for customization, with fine-tuning on their Tinker platform. I prefer that framing to the usual inflated launch language.
There is also a practical strategic angle here: if open-weight US models become viable alternatives to closed APIs, buyers get a little more negotiating power. And if pricing or terms shift later, teams that already know how to work with open models, custom MCP servers, or self-hosted inference are in a better position than teams that only built against one vendor endpoint.
What would make me cautious?
The model card and training data documentation sound far too short for a release of this size. The source highlights that the data documentation is basically limited to "public domain", publicly available internet content, publicly accessible repositories, and some third-party datasets.
That is not enough detail.
For hobby use, many people will shrug and move on. For enterprise use, especially anything customer-facing or regulated, thin documentation becomes an adoption tax. Legal, compliance, and procurement teams will ask obvious questions about provenance, filtering, retention, and evaluation. If the answers are vague, the burden shifts to the customer.
That is the second-order issue with releases like this: permissive licensing removes one barrier, but weak documentation creates another.
The API example is simple, which is good
The source includes a straightforward OpenAI-style chat completion call:
curl "https://tinker.thinkingmachines.dev/services/tinker-prod/oai/api/v1/chat/completions" \
-H "Authorization: Bearer $TINKER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "thinkingmachines/Inkling",
"messages": [
{"role": "user", "content": "Generate an SVG of a pelican riding a bicycle"}
],
"stream": false
}'
That OpenAI-compatible shape lowers the switching cost. If you already automate model calls in AI workflow automation, trying Inkling is unlikely to require a major client rewrite.
My take
This is a good release, and more importantly, it is a useful kind of good. Not flashy, not obviously state of the art, but practical: multimodal, permissive, and aimed at customization.
What I would watch is whether the surrounding ecosystem matures fast enough: better documentation, clearer evals, and realistic cost/performance comparisons against other open models. If that happens, Inkling could become the kind of base model teams actually build on. If not, it risks becoming another technically impressive model that enterprises hesitate to touch.
