Analysis · AI architecture and governance

Do you know where your models come from? Model provenance after the House report on Chinese open-weight models

Two House committees have recommended limiting Chinese open-weight models in federal work and creating a model provenance standard, and Cursor told them it would explore phasing such models out. Whatever you think of the politics, the question behind it is practical and global: can you say which models, and which models underneath them, power your products? Most companies cannot.

AnalysisSumanta Mukhopadhyay8 min read6 sources
#news#ai#governance#llm
ANALYSIS · ANALYSIS · AI ARCHITECTURE AND GOVERNANCE22 Nov 2026 · 21:20 GST

What happened

The House Select Committee on China and the House Homeland Security Committee published a joint report on Chinese open-weight models such as DeepSeek, Qwen and Kimi. It found that cost, customisation and near-frontier performance drive their adoption, and recommended restricting them in federal agencies and contract work, a national American open-weight model initiative and a model provenance standard1. Cursor told the committees it would explore phasing out Chinese-developed open-weight models because of geopolitical risk; CNBC reports its Composer coding model was built on a Moonshot Kimi model2.

The same week, India's Sarvam began offering hosted access to Chinese models including DeepSeek alongside its own, The Next Web reports, and the report itself warns that forcing companies off these models before trusted alternatives exist would raise costs3. The recommendations are not law. They are, however, a clear signal of the questions customers and regulators will start asking.

The real problem is hidden lineage

Cursor's case shows why this is harder than it looks. A product can be marketed under its own model name while being built on someone else's base model. Fine-tunes, distillations, merged models and hosted APIs all hide where the underlying weights came from. Software teams learned this with libraries: you depend not just on what you import but on everything it imports.

Illustration of an application depending on dependencies A, B and C, where A depends on D and E, and C depends on D and F
Transitive dependencies: models work the same way, with fine-tunes and distillations hiding their base. Source: CNCF TAG Security, Software Supply Chain Best Practices · CC BY 4.0

The CNCF's supply chain guidance describes exactly this structure for software, with direct and transitive dependencies each needing to be known and verified6. For models, the chain runs from base model to fine-tune to the hosted endpoint your product calls, and sometimes through a vendor that does not tell you.

What a useful model inventory records

Where used
Product featurescustomer-facingInternal toolscoding, support, analyticsVendor productsmodels inside SaaS you buy
The model
Name and versionas calledHostself-hosted, cloud, APILicenceand its restrictions
Lineage
Base model and developerand country of originModificationsfine-tune, distil, mergeTraining data noteswhere known
Controls
Evaluation resultsyour tasksApproved alternativestested fallbacksOwnerwho decides
Lineage is the layer most inventories miss, and the one customers now ask about.

Standards are starting to catch up. India's CERT-In expanded its bill of materials guidelines in 2025 to cover AI bills of materials alongside software, cryptographic and hardware ones4, and research such as AIBoMGen generates AI bills of materials during training, recording datasets, base model and configuration with signed attestations5. You do not need to wait for a final standard to keep a spreadsheet with the fields above.

Keep the right to switch

Provenance only helps if you can act on it. The architecture that makes switching cheap is the same one I recommend for cost and resilience: a model gateway between applications and models, evaluation sets for every important task, and at least one tested alternative for each.

  1. InventoryList every model and vendor model, with lineage where known
  2. AskRequest base model and origin from each AI vendor
  3. ClassifyFlag uses where origin matters to customers or rules
  4. PreparePut those uses behind a gateway with a tested alternative
  5. DecideChoose per use case, with cost and quality evidence
  6. RecordKeep the decision and its reasons for customers and auditors
Know, prepare, then decide: switching is a configuration change, not a rewrite.
SituationSensible stance
Selling to US federal agencies or their contractorsExpect origin questions now; prepare alternatives for affected uses
Regulated sectors such as finance and healthDocument lineage and evaluation; be ready to explain choices
Internal tools and experimentsRecord lineage; choose on quality, cost and data handling
Self-hosted open weightsOrigin matters less for data flow, more for licence and customer perception

Self-hosting an open-weight model keeps your data in your environment, whatever its origin. That is a different risk from sending data to a foreign API, and it deserves a different answer.

The CTO's take

This is not only an American question. Indian, European and Gulf customers ask their own versions, about sovereignty, data residency or licensing. The companies that handle it well will be the ones that already know what is inside their products and have kept the freedom to change it. That is an engineering discipline, not a political position. Related reading: per-market AI models, model routing, self-hosting open weights and evaluation sets.

If you need a clear view of the models inside your products and a plan to keep your options open, I am a fractional CTO and AI architect based in India, the United States and the UAE, working globally. You can analyse your challenge and bring the result to a call.

References
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
FAQ

Common questions

What is model provenance?

A record of where a model came from: who trained it, which base model it was derived from, how it was modified, where its weights are hosted and under what licence, so you can answer questions about origin, risk and obligations.

Why do derivatives matter?

Many products use models fine-tuned or distilled from a base model built by someone else. A product can depend on a model origin without naming it anywhere, in the same way software depends on transitive libraries.

Should companies stop using Chinese open-weight models?

That depends on customers, sector and jurisdiction. The House report's recommendations are not law. What every company should do is know where its models come from and be able to switch if customers or rules require it.

Keep reading

Related

Open to advisory, fractional CTO & select builds

Let's talk about what you're building.

The fastest route is a free 30-minute call. Prefer to write? Send a message and I'll reply personally within two business days.

Start a conversation

How would you like to reach me?

Book a 30-min callFastest. Pick any open slot on my calendar.→ Write in GmailOpens a pre-filled draft in your browser.→ Use my email appOutlook, Apple Mail or your default client.→

I reply personally within two business days. Time zones from the Americas and Europe to the Middle East and Asia-Pacific.