Two Open-Weight Launches, One Week, Opposite Terms: Read the Licence File
This week Inkling shipped Apache 2.0 weights on day one and Kimi K3 shipped an API with a promise of weights by 27 July. How a procurement reviewer reads each, why "open-weight" in a launch post is not a licence, and what the hardware floor does to a sovereignty claim.
This week two labs shipped a frontier-scale model under the word "open". On 15 July Thinking Machines Lab released Inkling, with weights you could download that afternoon. On 16 July Moonshot AI announced Kimi K3 as an open-weight model, with no weights and no licence text to go with it. Same word, same week, opposite terms.
"Open weights" is a marketing category. The licence file is the fact, and a procurement reviewer only ever reads the second one.
What was actually on disk by Friday
Inkling came first. Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati, published it on 15 July: a mixture-of-experts model with 975B total parameters and 41B active per token, 66 layers, six of 256 experts routed per token plus two shared. It takes text, images and audio in and writes text out. The context window runs up to 1M tokens and the training run covered 45 trillion tokens. The Hugging Face log shows a commit titled "Model release" at 17:58 UTC that day, with BF16 weights and an NVFP4 checkpoint for Blackwell hardware.
Kimi K3 came a day later and was bigger on every axis: 2.8T total parameters, a context window of 1,048,576 tokens, native vision. What shipped on 16 July was an API, the Kimi app and Kimi Code. The model ID was kimi-k3, priced at $3.00 per million input tokens, $0.30 cached and $15.00 output. On the weights, Moonshot's own post was precise: "The full model weights will be released by July 27, 2026."
So by Friday an architect could hold one of these two models inside their own perimeter. The other is a hosted API from a Chinese lab with a delivery date attached.
Inkling: Apache 2.0, with two footnotes
Inkling is the easy case, and it still takes a reviewer three steps.
- The declarationThe model card metadata reads license: apache-2.0, with a licence link to apache.org. No revenue cap, no user cap, no attribution duty. It cleared an Apache-2.0 / MIT floor on the day it shipped.
- The missing fileThe repository has no LICENSE file. The grant lives in the card metadata and nowhere else.
- The second documentA separate Model Acceptable Use Policy, effective 15 July 2026, says: "By accessing, downloading, or using any Model Materials, you agree to be bound by this Model AUP." It is a standard prohibited-use list with no revenue threshold, no user threshold and no geographic restriction. It never mentions Apache 2.0.
None of that makes Inkling a bad choice. It makes it a file that counsel has to look at once. Apache 2.0 carries no field-of-use limits of its own, so how a separate policy that claims to bind every downloader sits beside it is a legal question, and we do not know the answer.
What a reviewer can do without counsel is fix the evidence. Save the card metadata and the policy text as they stood on the day of download, next to the checkpoint hash. When the grant lives on a web page instead of in a file that travels with the weights, that archive is your proof of what you accepted.
On capability the vendor was unusually plain. Thinking Machines wrote that "Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization." Its own table agrees. The figures are vendor-reported, with comparison scores generated on 14 July: 54.3 on SWE-bench Pro against 58.6 for Kimi K2.6, 55.4 for DeepSeek V4 Pro and 80.0 for Claude Fable 5, and 74.1 on MCP Atlas against 83.3 for Fable 5. A base for fine-tuning, and honest about it.
Sovereign on paper: what can be checked this week
The case for K3 in an EU architecture will be sovereignty: frontier-class weights inside your own perimeter instead of a call to api.moonshot.ai. This week that case rests on a promise, and the one hard number Moonshot has published already sets a floor under it.
Moonshot's launch post recommends "supernode configurations with 64 or more accelerators". At 2.8T parameters that is not a mid-market server room. For most EU buyers "self-hosted K3" will mean K3 on somebody else's EU hardware, and a host can only offer it once the weights and a licence exist. Inkling gives the comparison: 41B of 975B parameters active per token, roughly 952 GB in BF16, with an NVFP4 checkpoint already on Hugging Face. That still means a multi-GPU Blackwell node, but it is a node you could order today.
Three things about K3 cannot be checked until the 27 July delivery: the licence text, the active parameter count, and which hosts will serve it inside the EU. Each of them decides whether "open" means anything for your perimeter. None of them is in the launch post.
Our read
The word "open" told a buyer nothing that mattered this week. One model came with a licence you can name in a procurement policy, plus two loose ends. The other came with a press cycle and a delivery date. Neither fact could be read off the headline.
What we would check before either model goes near a router:
- Weights on disk, licence file read. Promotion is gated on the artefact, never on the launch post. Until 27 July at the earliest, routing to K3 means sending data to Moonshot's hosted API and nothing else.
- Say which floor you mean. Our own floor for anything deployed on-prem is Apache 2.0 or MIT. Inkling meets it on the metadata. K3 cannot be judged yet, because "open-weight" in an announcement is not a licence. Write the sentence so a reviewer can apply it without calling you.
- Ask the host in writing. When a third party offers to serve K3 for you, its licence position becomes part of your supply chain.
- Count the accelerators before the word sovereign goes on a slide. Moonshot's own recommendation is 64 or more.
- Run your own eval set. Both vendors' tables are vendor-reported, and Moonshot ran K3's numbers on its own Kimi Code harness.
Two things we still do not know: how Thinking Machines' policy interacts with an Apache 2.0 grant, and what Moonshot's licence will say. Both belong in the file before a contract does.
Licence review, the routing policy and the eval gate between a model card and production are the unglamorous part of what Sebrona builds. If you are weighing an open-weight model for an EU deployment and want a second reader on the licence and the hosting map, write to info@sebrona.com.
Reading
Every figure and quote above comes from one of these. Licence texts were read from the raw files, not from coverage.
- Introducing InklingRelease date, 45 trillion training tokens, context options, NVFP4 checkpoint, hosting partners, the Inkling-Small preview.
- thinkingmachines/InklingParameter counts, architecture, the apache-2.0 card metadata, the absent LICENSE file, the vendor-reported benchmark table and the commit log.
- Model Acceptable Use PolicyEffective date and the clause binding anyone who downloads the model materials.
- Kimi K3 tech blogThe 27 July weights promise, the recommendation of 64 or more accelerators, and the note that K3's scores were run on the Kimi Code harness.
- Kimi API pricingModel ID and the $3.00 / $0.30 / $15.00 price line.