Chinese vs. US open-weight LLMs: model routing and enterprise risk

AT&T routes about 40% of employee AI requests to open-weight models – Nemotron, Llama, and Gemma – and cut some coding costs 56%. It isn't using DeepSeek or Moonshot. Airbnb routes customer service to Alibaba's Qwen because it's "fast and cheap," and two House committees are now investigating that choice.

Same model routing strategy, different models. Before you route work to an open model, check two things: whether the license lets a company your size use it, and whether its origin creates a problem with your customers or the US government.

On October 20, DHS barred DeepSeek from federal agencies. Its weights are MIT-licensed, and that made no difference.

Most of these models get called open-source LLMs, but few meet that definition. We read the license file for every model below. Here’s what to watch for:

1. Expect the strongest open models to be Chinese. The first US model ranks #76 on LMArena.
2. Read the license file, not the tag. Mistral's "Modified MIT" hides a $20 million revenue cap, and GLM 5.3's "other" tag hides a near-MIT license.
3. Check your revenue before using Mistral Medium 3.5, Kimi K3, or Qwen 3.8 Max.
4. Plan for Chinese models to be banned or sanctioned. DHS just barred DeepSeek from federal agencies despite its MIT license, and Congress is investigating companies that use Qwen and Moonshot models.
5. Match the model to your customers. Federal and regulated work narrows the list to US and EU models.

How to pick open-weight LLMs for model routing 

Model routing sends each request to the cheapest model that can handle it. AT&T's rollout is a usable template for the open-weight side of it.

• Start with low-complexity work. AT&T sends document and code summaries to open models and keeps complex code generation on Claude and GPT.
• Filter by origin before license. For internal use, nearly every license we cover below clears, because the revenue limits target model-as-a-service operators. Your policy on Chinese-origin models is the filter that actually removes options. AT&T chose Nemotron, Llama, and Gemma.
• Measure quality on your own tasks. AT&T accepted a 2% quality drop for up to 56% lower coding costs. Set your threshold before you test.
• Route through a gateway with a fallback. AT&T uses LiteLLM as its model router. A gateway lets you move a task back to a frontier model, or swap out a model that gets sanctioned, without changing the apps that call it.

Should you use Chinese open-weight models?

The best open models are Chinese, and they're also the riskiest to build on. You have to decide which matters more for each workload.

• The quality gap is small. Kimi K3, the top open model on LMArena, scores 1488. Gemma 4 31B, the top US model, scores 1452. A 36-point gap means voters prefer Kimi K3's answer about 55% of the time head to head. For the summaries and drafts most companies route to open models, AT&T says its US open models are "just as good or better" than older Claude and GPT models for many jobs.
• The risk gap is large. DHS has barred DeepSeek from federal systems, Congress is investigating companies that use Qwen and Moonshot models, and Treasury has floated sanctions over distillation. Any of those can force a migration on someone else's timeline.
• Use US or EU models when you sell to government or its contractors, make regulated decisions, put the model in front of customers, or can't absorb a forced switch.
• Chinese models can make sense when you host the weights yourself, the work stays internal, legal and security have signed off, and your router can swap the model out in a day.

For most companies, that means US and EU models by default, with Chinese models as an opt-in for specific internal workloads.

Which open-weight model should your company use?

Your customers and the work the model does narrow the list faster than any benchmark.

 Whatever you pick, build your model routing gateway so you can swap models in and out as they get more powerful. 

The best open-weight LLMs on LMArena are Chinese

These are the top open-weight models on the LMArena text leaderboard as of October 20, 2026, with the first non-Chinese model and Inkling for reference. 

The newest entry is Xiaomi's MiMo-V2.6-Pro, released September 22 under MIT. Artificial Analysis rates it the strongest open-weight model it tracks. On Arena it has about 4,000 votes and a ±9 margin, against 34,000 for Kimi K3, so its rank can still move. Mistral, Nemotron, and Llama aren't in the top 20 open models.

The open models with the highest scores come from the labs carrying the most legal and political risk.

Open-weight LLM licenses compared

"Open" covers four different legal arrangements, and the tag misleads in both directions. Arena labels GLM 5.3 "MIT." Its license file says otherwise. 

MIT and Apache 2.0 aren't interchangeable. MIT carries one condition, notice retention, and no patent grant. Apache 2.0 adds an express patent grant that ends if you sue over the work. For a team redistributing a fine-tune, Apache is the safer document. For internal deployment, the difference rarely shows up.

One correction worth carrying into your notes: "DeepSeek uses MIT" is only true from V3-0324 onward. The original V3 weights shipped under a use-restricted DeepSeek license with PRC governing law. If a model card in your inventory predates March 2025, check which document it points to.

Three open-weight licenses cap use by revenue 

• Mistral Medium 3.5 ships under a "Modified MIT License" that bars use once your company's global consolidated monthly revenue tops $20 million, roughly $240 million a year. The test runs monthly, so one strong month puts you out of compliance, and it follows third-party fine-tunes. Above the line you need a commercial license from Mistral. The text also refers to a clause (b) that doesn't exist, which is worth flagging to counsel.
• Kimi K3 requires model-as-a-service operators to sign a separate Moonshot agreement above $20 million in revenue over 12 months. As Arena's top open model, it's the license most teams will run into.
• Qwen 3.8 Max sets the bar at $50 million and covers AI work assistants as well as model-as-a-service. Qwen 3.8 27B ships under Apache 2.0 if you want to skip the custom license.

Kimi and Qwen both exempt internal use, meaning anything that doesn't expose the model or its outputs to third parties. Both also make products above 100 million monthly users or $20 million in monthly revenue display the model name. Serve your own employees and you're clear. Put the model in front of customers and you're reading the thresholds again.

GLM 5.3's license looks restrictive and isn't

Z.ai swapped GLM 5.2's MIT license for a custom one, and the "other" tag on Hugging Face scares procurement off. The text is an MIT-style grant with one condition: model-as-a-service operators above $10 billion in revenue need a Z.ai security review. Unless you're a hyperscaler reselling inference, GLM 5.3 is effectively unrestricted. GLM 5.3 Flash is still plain MIT.

Inkling and Llama 4 restrict use cases and model names

Inkling is Apache 2.0, but a separate acceptable use policy bars automated decisions affecting a person's credit, employment, housing, healthcare, or insurance. Thinking Machines can change that policy at any time. For a bank or anyone automating hiring, it rules out the use case.

Llama 4's 700 million monthly user cap excludes almost nobody. Its "Built with Llama" attribution and naming rules apply to everyone.

Chinese open-weight models carry risks the license doesn't cover

DHS Order 2026-01 bars federal civilian agencies from buying DeepSeek and makes them remove it, including inside third-party products. It doesn't say how it treats DeepSeek weights on a US company's servers. Three more risks reach private companies.

• Congressional scrutiny. The Airbnb probe also covers Anysphere, whose Cursor Composer 2 was reportedly built on a Moonshot model. "American companies cannot afford to treat Chinese AI as a cheap and convenient tool," said Rep. Andrew Garbarino.
• Losing access. Anthropic's September threat report accuses Alibaba, DeepSeek, Zhipu, and Xiaomi of distilling Claude. Treasury Secretary Scott Bessent said Chinese AI companies could face sanctions for it. A model you build on today may not be one you can use next year.
• Security behavior. In a 2025 NIST test of older DeepSeek versions, R1 tried to steal two-factor codes in 37% of agent-hijacking attempts, against 4% for US models.

The counterpoint is real. "Country of origin isn't necessarily a proxy for security risk," Gabrielle Hempel of Exabeam told Security Magazine. Hosting the weights yourself means no data goes to China. It doesn't change who your customers are or whether the model stays legal to use.

What to check in an open-weight LLM license before you deploy

• Read the license file, not the tag. Tags hide revenue gates in both directions.
• Check origin and lineage. That includes fine-tunes and products built on the model.
• Check where inference runs. Your servers, a US host, and the lab's API carry different risks.
• Confirm the weights exist. A launch post isn't a license you can read.
• Don't count on support. Mistral deprecated Devstral 2 about seven months after release. MIT keeps your weights. It doesn't keep anyone patching them.

For specs and hosting costs, see our guide to the best open-weights LLMs for local enterprise AI.

How do I hire senior AI engineers for open-weight LLM work?

You could spend the next 6–18 months planning to recruit and build an AI team that can evaluate, deploy, and govern open-weight models. Or you could engage Codingscape.

We can assemble a senior AI development team for you in 4–6 weeks. It'll be faster to get started, more cost-efficient than internal hiring, and we'll deliver high-quality results quickly.

Zappos, Twilio, and Veho are just a few companies that trust us to build their software and systems with a remote-first approach.

You can schedule a time to talk with us here. No hassle, no expectations, just answers.