THE VERDICT
● For capability, the Chinese open-weight models lead on several benchmarks and are cheaper per token.
● For procurement, Laguna S 2.1 and Gemma 4 are easier to defend and always were.
● The advisory changed nothing technical. It changed what you have to write down.
Where each sits
| Model | Origin | Licence | Price / 1M |
|---|
| Kimi K3 | Moonshot, CN | Modified MIT | $3.00 |
| GLM-5.3-Flash | Z.ai, CN | MIT | $0.15 |
| Tencent Hy4 | Tencent, CN | Apache 2.0 | Self-host |
| DeepSeek V4 Flash | DeepSeek, CN | MIT | $0.14 |
| Laguna S 2.1 | Poolside, US | OpenMDW | $0.10 |
| Gemma 4 | Google, US | Apache 2.0 | Free |
| K2 Horizon | IFM | Plus training data | Free |
WHAT THE ADVISORY DID AND DID NOT DO
It did not ban anything, restrict anything, or make any claim about these models being unsafe to run.
What it did is convert a circulating allegation into a documented position from three US agencies — which is the form a procurement team has to record and answer for.
Which one
| If you are... | Pick |
|---|
| A startup shipping a product | Whatever performs best. Nothing prevents it and nothing changed technically |
| Selling into government or defence | Western origin. This will appear in a questionnaire |
| In a regulated sector | K2 Horizon. Published training data answers the question others cannot |
| Cost-constrained above all | Laguna S 2.1 at $0.10 is Western and the cheapest here |
| Already deployed on a Chinese model | Nothing requires a change. Configure a fallback and document what you use |
FAQ
Are Chinese open-weight models banned now?
No. The NSA, CISA and FBI issued an advisory, which is guidance rather than a prohibition.
Is Laguna as good as Kimi K3?
Different strengths. Kimi leads on agentic coding benchmarks; Laguna posts strong Terminal-Bench results at a fraction of the price. Test both on your own work.
Does self-hosting change anything?
Weights on your hardware are unaffected operationally. The allegation concerns how the models were built, not what they do when you run them.