Scroll through this week's open-model leaderboards on X and one pattern is hard to miss. A tally by Dr John Seach counts at least eight Chinese labs that between them shipped more competitive open-weight models in 2026 than the rest of the world. DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Zhipu's GLM: the same names sit at the top of nearly every open ranking.
For a European company weighing local AI, that raises an awkward question. Can I deploy a model trained in China? The short answer: it depends on where you run it, not on who trained it.
Who leads open weights in 2026
By the rankings now circulating, four of the top five open-weight models come from Chinese labs, with Meta's Llama the notable outlier. Community reports put these models within a few points of the best closed systems such as GPT-5.5 and Claude Opus 4.8 on coding and reasoning, at a fraction of the cost. The exact gap is hard to verify, but the direction holds across several independent leaderboards.
A current example: Moonshot's Kimi K3. We covered the planned release ahead of the date (see Kimi K3: What SMBs need to know). What is new since that piece is that the weights actually dropped on July 27. According to Times of AI on X, K3 lands around fourth on the Agent Arena leaderboard, level with leading closed models. Read numbers like that as a snapshot, not a promise.
The part that matters in practice is the license. DeepSeek, for instance, ships under the permissive MIT license, others under Apache 2.0. Both allow commercial use and local fine-tuning without asking the vendor.
The reflex worry, and what is fair about it
Being wary of a model trained in China is not silly. It just often aims at the wrong target. Two questions are genuine:
- Where does my data go? Central for a hosted service. With static weights you run yourself, it evaporates.
- What is in the training data? Relevant for regulated output, say when a model drafts legal or medical text. That question stands regardless of origin, and applies to Western models just as much.
What is rarely a technical problem: the vague notion that a downloaded model might quietly phone home during use. A set of weights is a file. It computes when you tell it to, and does nothing else.
Why local deployment dissolves the origin question
This is exactly where the approach we run with clients kicks in. An open-weight model running on your own hardware is a closed system. No request leaves the network and no vendor sees your prompts. Whether the weights were made in Hangzhou, Paris, or Palo Alto changes none of that, because the traffic stops at your firewall.
That is the heart of our take on data sovereignty: sovereignty is not about whose model you use, it is about where the processing happens. An air-gapped DeepSeek or Qwen exposes confidential files less than a closed EU cloud API you have to trust the operator on. What such an environment looks like in practice we describe under local AI.
There is a pleasant side effect for verification. Because the weights are open, you can inspect them, test them against your own cases, and version them deterministically. A closed endpoint gives you none of that.
What that means for hardware
One practical point often decides more than the ranking: does the model fit on your machine? As a rough community rule, a dense model in the 27B to 32B range at 4-bit quantization gets by on roughly 20 to 24 GB of memory, a 70B model closer to 40 to 48 GB. That puts most day-to-day work on a single workstation or a Mac Studio with generous unified memory.
The absolute flagships from many Chinese labs, DeepSeek and Kimi among them, are large mixture-of-experts systems weighing hundreds of gigabytes in memory. Those need either a machine with a lot of RAM or a smaller distilled variant. For most SMBs the biggest model is not the right pick anyway. The smallest one that solves the task reliably is, because it saves hardware and power while keeping response times usable.
What you still need to check
Local does not mean carefree. Before a Chinese open-weight model moves into a business process, a few sober checks pay off:
- Read the license text, not the label. MIT and Apache 2.0 are uncritical. Some models carry bespoke community licenses with usage clauses or revenue caps. That difference decides whether you may use the model commercially at all.
- Reproduce the benchmarks yourself. Do not lean on the vendor's numbers. A small test set built around your real tasks tells you more than any leaderboard. Based on our reading, some models perform very differently in your language or domain than in the English benchmark.
- Review output where it counts. Regulated content needs a human control step anyway, whatever the model. That is a question of process, not provenance.
Clear those three and you are deciding on capability and license, not on a gut feeling about a flag.
A European alternative remains
None of this makes Chinese models mandatory. France's Mistral also ships competitive open weights, with the advantage of a European legal chain and strong performance in European languages. We covered it in detail (see Mistral Open-Weight: Sovereign Local AI in Europe). The practical upshot is the same: choose on capability and license, then run the result locally.
If you want to know which open model fits your tasks, your hardware, and your compliance needs, we set up a scoped pilot project and test the candidates against your real data before you commit.