Inkling and the Return of US Open-Weights Releases

For most of 2025 and 2026, the open-weights frontier had a geography problem. The most capable downloadable models came overwhelmingly from Chinese labs: Alibaba’s Qwen, DeepSeek, Zhipu’s GLM, Moonshot’s Kimi, MiniMax, Tencent’s Hunyuan. US labs had mostly retreated to closed APIs, with OpenAI’s gpt-oss and Google’s Gemma as the notable exceptions and Meta’s Llama sitting under a community license rather than a standard open one.

In July 2026, Thinking Machines Lab released Inkling, its first open-weights model: a large mixture-of-experts design with roughly 975 billion total parameters and about 41 billion active per query. Whatever its benchmark position turns out to be, it matters as a market signal.

Why Geography Matters Here At All

It should not matter, in an ideal world. Weights are weights, licenses are licenses, and a good model is a good model regardless of the passport of the team that trained it.

In practice it matters for one specific reason: procurement.

A European law firm, a US healthcare provider, or a government contractor evaluating AI tooling has to answer questions about data flow, jurisdiction, and supply chain. When they use hosted inference, the answers depend on who runs the servers. When the only credible open models come from one region, organizations with restrictions on that region have fewer options, and “fewer options” in practice means “use a closed US API and stop asking.”

A US open-weights release at frontier scale widens the set of answers available. It does not make anyone’s model better; it makes some procurement conversations shorter. We wrote about how those conversations go in how to audit an AI vendor in 2026.

The Architecture Detail Worth Noticing

975 billion total parameters with 41 billion active is a sparse design: only a fraction of the network computes for any given token. That gets you compute costs closer to a 40B model with capability closer to something much larger.

What it does not get you is a model you can run at home. Memory is set by total parameters, not active ones, and holding 975 billion parameters is datacenter work regardless of how few of them fire. This is the point that trips up almost everyone reading a mixture-of-experts specification for the first time, and it deserves repeating whenever a large open release lands: active parameters predict speed, total parameters predict whether it loads.

So Inkling, like Kimi K3 and DeepSeek V4, is open in the sense that anyone can inspect, fine-tune, or serve it, and closed in the practical sense that you will reach it through somebody’s API. We ran the numbers in can you run a trillion-parameter model locally.

What Frontier Open Releases Do For Small Models

The mechanism by which a 975B model eventually helps your laptop is indirect but reliable.

Large open models are raw material. They are used to generate training data, to distill into smaller students, and to establish techniques that get reimplemented at smaller scale. The 1B to 8B models that actually run on consumer hardware improve generation over generation partly because there is a richer supply of large open models to learn from.

That lag is roughly measured in months, and it is the reason a current 4B model outperforms a 13B model from two years ago. If you use a small local model, frontier open releases are good news arriving later, in a package you can actually run.

What It Does Not Change

Three things stay exactly where they were.

Hosted open weights has cloud privacy properties. Running any large open model through a provider’s API means your text is on their servers under their retention policy. The license of the weights is irrelevant to what the host logs. This is worth stating every time, because “open source model” gets read as “safe to paste anything into.”

Frontier scale is still wrong for rewriting. The facts are already in the text you paste; nothing needs to be recalled or derived. Extra capability shows up as extra initiative, and initiative is a defect when the task is a tone change. See why small models beat big ones for rewriting.

Your instruction still matters more than your model. The gap between a vague instruction and a precise one exceeds the gap between model tiers on this kind of work.

Keeping Perspective On Release Weeks

July 2026 alone brought Kimi K3, Inkling, Tencent’s Hunyuan 3.0, an NVIDIA Nemotron variant, and Claude Opus 5, with GPT-5.6 landing publicly earlier in the month. Trying to evaluate each one is a full-time job that produces no output.

A stance that survives contact with this pace: treat frontier releases as market news and re-evaluate your own setup on a schedule rather than on announcements. The questions that decide your experience are local. Can I run something good enough on hardware I own? Is the license clean? Does my instruction hold the model to the task? None of those change when a lab ships something enormous.

There is a fuller method in how to keep up with AI model releases.

Common Questions

Can I download and run Inkling?

You can download it. Running it requires multi-GPU datacenter hardware, so in practice you will use it through a hosting provider if you use it at all.

Is it better than the Chinese frontier open models?

Benchmark comparisons at this scale shift with every release and rarely predict performance on a specific task. If you have a real use case, test it on that rather than reading tables.

Does a US-developed model mean my data stays in the US?

No. Where the model was trained has nothing to do with where inference runs. Data residency depends on your hosting provider, and it is a question to ask them in writing.

What should I actually run for confidential writing?

A small open-weights model on your own machine, under Apache 2.0 or MIT. Slightly less polished than a frontier model, and the only configuration where your draft never leaves your hardware.

Download Wrivio for Windows to run a small open-weights model in-process, unaffected by whatever ships next week.