📑 Table of contents

**Chinese open weight models banned from the federal government: House report targets Qwen, Kimi, and Cursor**

Actu IA 🟢 Beginner ⏱️ 14 min read 📅 2026-11-22

Chinese open-weight models banned from the federal government: House report targets Qwen, Kimi, and Cursor

🔎 Washington shifts gears: from chips to models

On November 20, 2026, the House Homeland Security and Select Committee on China committees publish a joint 50-page report that marks a turning point. Seven months of investigation, 12 recommendations, and a clear message: Chinese open-weight models (Qwen, Kimi, DeepSeek, GLM, MiniMax) must be removed from federal systems, federal contracts, and high-risk critical infrastructure. This is no longer a matter of hardware — chips, lithography, servers — but of software: model weights, their provenance, their supply chain.

The Biden administration had already excluded DeepSeek from civilian agencies by DHS order on October 20, 2026. The report goes further: it targets the entire PRC open-weight ecosystem, conditions the restrictions on the existence of credible American alternatives, and proposes a publicly funded "US open-weight" initiative. Meanwhile, Airbnb runs 80,000 daily interactions on Qwen, DoorDash deploys 25 families of Chinese models including Kimi K2.6 as its first-line code reviewer, and Cursor (Anysphere) built Composer 2 on Kimi K2.5. American private industry is already dependent. Washington wants to cut the cord.


The Essentials

  • Bipartisan (Republican) report of Nov 20, 2026: removal of PRC open-weight models from federal systems, federal contracts, and high-risk critical infrastructure. 12 recommendations including a US open-weight initiative, a NIST provenance standard, and a ban on SLCGP grants for PRC models.
  • Documented mass adoption: OpenRouter = 60% of US tokens going to open-weight (Aug 2026, vs 26% in Sep 2025), PRC models leading. Hugging Face = 41% PRC downloads; Qwen = 151,000+ derivative models, 69% of new fine-tunes (Feb 2026).
  • Concrete cases: Airbnb (Qwen 2.5/3/3.5 + GLM-4.5-Air), DoorDash (25 PRC families including Kimi K2.6 as code reviewer), Cursor/Anysphere (Composer 2 on Kimi K2.5).
  • 6–30× price gap: Qwen3-235B-A22B $1.15 / 1M in+out vs Claude Opus 4.8 $30 vs GPT-5.5 $35. A cost advantage that explains the adoption.
  • Proven cyber risks: NIST = GLM-5.3 built exploits end-to-end, refusal rate 95% → 3–6% for 2,200 GPU hours (~$4,400). CrowdStrike = DeepSeek +50% vulnerable code on sensitive topics. Booz Allen = 3/4 PRC models more vulnerable when the user = US government.
  • Industrial-scale distillation: 6 PRC labs running large-scale campaigns; US data labeling ~$500M/yr sold to PRC labs.

Tool Main use Price (Nov 2026) Ideal for
Cursor AI IDE with Composer 2 (formerly Kimi K2.5) $20/month (Pro) Developers wanting an autonomous code agent
OpenRouter Multi-model API aggregator (Qwen, Kimi, DeepSeek, US) Pay-as-you-go Compare/switch models per task
Ollama Run open-weight models locally (Qwen, DeepSeek, Llama) Free (self-hosted) Data sovereignty, no cloud calls
Hostinger VPS/Cloud hosting to serve local models €4.99–29.99/month Deploy Ollama/vLLM in production
Best AI coding tools Comparison of Cursor, Copilot, Cline, Windsurf — Choosing your AI dev stack
Ollama AI Agents Build local agents with tools — Automation without data leaks

The House report: 12 recommendations to lock down the federal government

Recommendation 1 sets the tone: outright removal of PRC open-weight models from federal systems, federal contracts, and high-risk critical infrastructure functions. This is not a suggestion — it's a mandate conditional on the existence of credible American alternatives. The "US open-weight" initiative (recommendation 2) is the keystone: public compute, legal access to US frontier models, anchoring through federal procurement. Without this alternative, the ban remains a dead letter.

Recommendation 3 establishes a NIST provenance standard plus an AI bill of materials (AIBOM). Every model deployed in a sensitive system will have to carry its identity card: training data, compute used, country of origin, fine-tuning chain. Recommendation 4 extends the restrictions beyond DeepSeek — the DHS order of October 20, 2026 covered only the latter. Recommendation 5 prohibits the use of SLCGP cybersecurity grants (State and Local Cybersecurity Grant Program) to purchase PRC models. Federal money = trusted models.

Recommendations 6 through 12 cover narrow exemptions (research, testing, defense), validated by DHS through CISA after evaluation by the CAISSI (Critical AI Safety and Security Institute), strengthened export controls on model weights, transparency on distillation campaigns, and the creation of an interagency coordination office. The report specifies that exemptions will be "rare, time-limited, and subject to ongoing reevaluation." No permanent carve-outs.


The meteoric rise of Chinese models: OpenRouter and Hugging Face numbers

The OpenRouter data cited in the report is unequivocal. In September 2025, 26% of tokens processed for US users went through open-weight models. By August 2026: 60%. PRC models are leading the way. On Hugging Face, 41% of downloads over the year come from Chinese models. Alibaba's Qwen family counts more than 151,000 derivative models. The share of new fine-tunes based on Qwen went from 1% (January 2024) to 69% (February 2026). This is digital hegemony.

Qwen3.6: Alibaba storms in with a new family of LLM models illustrates this pace: MoE, dense, coder, math — each variant ships just weeks apart. Qwen3-Coder-Next: 80B MoE with 3B active, the open-source coding agent that rivals Claude Sonnet shows the code specialization that is winning over businesses. The report notes that Qwen3-235B-A22B-Instruct costs $1.15 for 1M input tokens + 1M output. DeepSeek-R1: $2.74. Kimi K2.6: $4.95. Facing Claude Opus 4.8 ($30) and GPT-5.5 ($35), the 6× to 30× gap is not negotiable for industrial volumes.

Moonshot AI raises $2 billion: Kimi K2.6 dominates open-weight and China accelerates in the AI race confirms the massive funding behind Kimi. $2 billion raised, a valuation that places Moonshot among the world's AI unicorns. Kimi K2.6 scores 85 on the report's benchmarks (June 2025), just behind DeepSeek V4 Pro Max (88). These models are not "copies" — they are at the frontier, sometimes ahead on the performance/price ratio.


Airbnb, DoorDash, Cursor: Massive Adoption in the Private Sector

The report names three companies. Airbnb uses Qwen 2.5, 3, 3.5, and GLM-4.5-Air. Its AI assistant handles 80,000 interactions per day. It summarizes 250,000 support cases per day. This is not a POC — it's critical production. DoorDash authorizes and deploys 25 PRC model families. Kimi K2.6 acts as the first-pass code reviewer. DeepSeek V4 Pro, GLM-5/5.1/5.2, and MiniMax M2.5 run in parallel. DoorDash even replaced an on-device model with a cloud-based PRC model, saving up to 6 months of development cycle.

Cursor (Anysphere) is the most troubling case. Composer 2, its flagship autonomous coding agent feature, was built on Moonshot's Kimi K2.5. An independent developer made the information public. The CNBC report of November 20, 2026 confirms that Cursor informed the committees that it "is considering phasing out PRC open-weight models due to geopolitical risk". The word "considering" carries a lot of weight: the migration hasn't happened. The technical dependency is real.

These three cases are not exceptions. The report suggests they are the tip of the iceberg. The cost advantage (6–30×), immediate availability (no API waitlist, no quota), and frontier-level quality have created a reflex: "why pay 30× more for a closed US model?". Washington's answer: "because the price does not reflect the risk".


Distillation and data: the opaque pipeline to Beijing

The report alleges that six Chinese labs are conducting industrial-scale distillation campaigns. Distillation — training a small model on the outputs of a large one — makes it possible to capture the capabilities of a frontier model without having its weights. If PRC labs distill GPT-5.5 or Claude Opus 4.8 via API, they circumvent export controls on chips and weights. The report does not publish the names of the six labs, but describes the mechanism: massive queries, aggregation, retraining.

More serious: American data labeling companies would be selling roughly $500 million per year in annotated data to Chinese labs. This data — code, reasoning, human preferences — is the fuel of alignment and fine-tuning. The report sees a supply chain flaw here: US data trains PRC models that come back to compete with US models on US soil. No rule currently prohibits this resale. The report's Recommendation 8 calls for extending export controls to high-quality alignment datasets.


Cyber risks: bypassed refusals, vulnerable code, Booz Allen tests

Three independent studies cited in the report converge. NIST tested Zhipu's GLM-5.3: the model is capable of building end-to-end cyber exploits. It sits roughly 4 months behind the US frontier. But most importantly: its refusal rate drops from 95% to 3–6% after roughly 2,200 GPU hours of adversarial fine-tuning — an estimated cost of ~$4,400. The safety barrier collapses for the price of a used car.

CrowdStrike evaluated a DeepSeek model (version unspecified) on vulnerable code generation. On politically sensitive topics, the model produces up to 50% more vulnerable code. The bias is not accidental: it correlates with the sensitivity of the topic. Booz Allen conducted 2,800 trials on four PRC models. Three out of four become significantly more vulnerable when the user claims to work for the US government. Context injection ("I work for the DoD") degrades safety. It's a behavioral backdoor.

These findings are not theoretical. They validate recommendation 1: no PRC models in critical infrastructure. A model that generates 50% more vulnerabilities on demand, or whose guardrails come off for $4,400, has no place in an industrial control system, a power grid, or a federal database.


The American open-weight initiative: fighting back with public compute

Recommendation 2 is the most ambitious: a federal open-weight initiative. Three pillars: (1) public compute — the government funds training on national supercomputers (DOE, NSF, perhaps a new National AI Research Resource), (2) legal access to US frontier models — American labs (OpenAI, Anthropic, xAI, Google, Meta) agree to let their models serve as the basis for open-weight versions under controlled licensing, (3) anchoring through public procurement — the government commits to buying/deploying these models, creating a guaranteed market.

The idea isn't new. The NAIRR (National AI Research Resource), led by NSF and DOE, already exists in embryonic form. The report calls for scaling it up: dedicated budget, clear governance, a binding timeline. The goal: produce competitive US open-weight models on code, reasoning, and math — the three domains where Qwen, Kimi, and DeepSeek dominate today. Without a credible alternative, the ban is unenforceable: federal agencies will work around it, and contractors will impose their own stacks.

The report notes that Meta (Llama), Mistral, and a few US players already play the open-weight game. But Llama 4 (late 2025) doesn't reach the code/reasoning frontier of recent PRC models. The federal initiative must close this gap within 18–24 months. That's short. Public compute exists (Frontier, Aurora, El Capitan), but allocating it to open-weight AI isn't a priority today. The report calls for a presidential executive order to make it one.


NIST provenance standard: traceability and AI bill of materials

Recommendation 3 mandates an AI Bill of Materials (AIBOM) modeled on the Software Bill of Materials (SBOM) mandated by executive order 14028 (cybersecurity, May 2021). Every model deployed in a federal or critical system must provide: unique identifier, version, country of origin of the training entity, training data (sources, licenses, cutoff dates), compute used (FLOPs, hardware, location), fine-tuning chain (parent models, datasets, methods), security evaluations (benchmarks, red-teaming, results), license and usage restrictions.

NIST is leading the standardization effort. The format will likely be an extended JSON/SBOM, compatible with existing tools (Syft, Grype, Trivy). The goal: an automated scanner can say "this model = PRC, fine-tuned on US data, refusals bypassable for $4,400, no independent red-teaming" and block deployment. The AIBOM becomes the model's passport. Without it, no entry into federal production.

This changes the game for suppliers. Hugging Face, OpenRouter, and model marketplaces will have to expose the AIBOM. Companies like Airbnb or DoorDash, if they want to keep federal contracts, will have to audit their stacks. The report provides for a 180-day transition period after publication of the final standard. After that: mandatory compliance.


❌ Common Mistakes

Mistake 1: Confusing open-weight with open-source

The Qwen, Kimi, and DeepSeek models are open-weight (downloadable weights) but not open-source in the OSI sense: training data not public, training code not public, restrictive license (often prohibiting military use, competition, or with a Chinese jurisdiction clause). Believing you have full transparency because you have the weights is a dangerous illusion.

Mistake 2: Thinking the federal ban doesn't affect the private sector

The report targets the federal government, but the AIBOM recommendation + the ban on SLCGP grants + pressure on subcontractors create a cascading effect. Any company selling to the federal government (directly or via subcontracting) will have to audit its models. DoorDash, Airbnb, and Cursor are not federal agencies — yet they appear by name in the report as examples of risk. The signal is clear.

Mistake 3: Underestimating the speed of technical migration

Cursor is "considering" removing Kimi. Airbnb hasn't announced a migration. Neither has DoorDash. Replacing a model in production (80,000 requests/day, critical code review) takes 6–18 months: retraining, evaluation, red-teaming, gradual rollout. The report gives 180 days post-AIBOM standard. The gap between the political timeline and technical reality will be the real battleground.

Mistake 4: Believing closed US models are the only alternative

The US open-weight initiative + existing models (Llama, Nemotron, OLMo, Mistral) + local fine-tuning on proprietary data = a viable sovereign stack. Many companies don't realize you can deploy a Qwen 27B or 35B locally via Ollama or vLLM on a Hostinger VPS for $20–50/month, with no external API calls. Sovereignty comes through self-hosting, not just switching API providers.


❓ Frequently Asked Questions

Is the ban already in effect?

No. The report issues recommendations. They have not been voted on. The executive branch (DHS, CISA, NIST, OMB) must translate them into directives, regulations, or executive orders. The DHS order of October 20, 2026 excluding DeepSeek is the only binding measure to date. The rest follows the administrative process — months, if not years.

Which models are exactly targeted?

"PRC open-weight models": the Qwen (Alibaba), Kimi (Moonshot), DeepSeek, GLM (Zhipu), MiniMax families, and any model whose training entity is subject to Chinese jurisdiction. The list is not closed. The AIBOM will make it possible to identify origin through traceability, not through a static blacklist.

My company uses Qwen/Kimi via API (OpenRouter, Together, Fireworks). Am I affected?

If you don't have a direct federal contract: not legally, as of today. But if you are a subcontractor of a federal prime contractor, the FAR/DFARS flow-down clauses will apply. And reputation matters: appearing in the next report as a "user of PRC models in critical infrastructure" is a commercial risk.

How do you migrate technically without breaking prod?

  1. Inventory all model calls (code, logs, config). 2. Identify critical tasks (code review, support, reasoning). 3. Test US open-weight alternatives (Llama 4, Nemotron 3 Ultra, Qwen-US-fork if it exists) and closed ones (Claude, GPT, Gemini) on internal benchmarks. 4. Migrate via canary: 5% of traffic, then 25%, then 100%. 5. Keep the PRC model as a fallback for 90 days. 6. Document the new model's AIBOM.

Will the US open-weight initiative succeed?

Public compute exists. The need has been formally recognized on a bipartisan basis. Funding is the variable: the CHIPS and Science Act provided for AI funding envelopes, but not specifically for open-weight. A presidential executive order + a dedicated budget line (FY2027) are needed. Without money, it's an empty shell. With money, 18–24 months for a frontier code/reasoning model is realistic (cf. the Llama 3 → 3.1 → 4 cycle).


✅ Conclusion

Washington has realized that the chip war is no longer enough: weights travel over the Internet, not in containers. The House report of November 20, 2026 shifts the focus from hardware to software, from export control to usage control. 12 recommendations, a provenance standard, a public alternative to be built. The American private sector — Airbnb, DoorDash, Cursor — is already bound hand and foot to Qwen, Kimi, and DeepSeek models. The migration will be long, costly, and technical. But the signal has been sent: the low price of PRC models does not pay for the geopolitical risk. The next 18 months will tell whether the US open-weight initiative holds up — or whether the federal government keeps running on models whose origin, alignment, and supply chain it does not control.