Open Weights Policy

Summary: Operational policy framework for open-weight AI models — defining what "open weights" means in regulatory contexts, distinguishing from open-source AI, establishing release/licensing/eval norms, and mapping policy levers (compute, procurement, export controls, liability) to desired outcomes (diffusion, competition, security, sovereignty).

Definitional Precision (Policy-Critical)

Term Definition Policy Relevance
Open Weights Model weights publicly downloadable; user can run, fine-tune, quantize, distill, deploy anywhere Regulations targeting "model access" must specify weight-level vs. API-level
Open Source AI (OSI OSAID) Weights + training code + training data (or detailed description) all under OSI-approved licenses Higher bar; few models meet it (OLMo, BLOOM); regulations should not conflate
Open License License permitting commercial use, modification, redistribution (Apache 2.0, MIT, Llama Community) Procurement rules can require open licenses; liability shields may attach
Weight Release Act of publishing weights (staged, gated, or open) Export controls, compute thresholds, notification rules may trigger at release
Derivative Model Fine-tuned, quantized, distilled, merged, or otherwise modified version Liability, licensing, export control treatment of derivatives unclear

Policy Objectives (Weighted by Coalition Emphasis)

Objective Weight Key Metrics Primary Levers
Economic Diffusion 30% Startup count using open models; API cost reduction; sector adoption breadth Compute credits, NAIRR, procurement preferences, tax incentives
Competition/Market Structure 25% Number of viable model providers; price/performance convergence; switching costs Antitrust guidance, API interoperability mandates, open model hosting support
National Security / Sovereignty 20% Allied sovereign deployments; defense/ic open model usage; supply chain resilience Export control reform, defense procurement, allied tech agreements
Safety via Transparency 15% Vulnerabilities found/fixed in open vs closed; red team participation; eval coverage Funded open red teaming, mandatory pre-release evals > threshold, disclosure norms
Innovation Velocity 10% Time from paper to production; derivative model count; citation/usage metrics Research funding, open dataset/eval investment, talent visas

Policy Levers by Government Function

1. Compute & Infrastructure (DOE, NSF, Commerce, OMB)

Lever Policy Action Target
National AI Research Resource (NAIRR) Fund GPU hours for open model training/fine-tuning; prioritize academia/startups Diffusion, talent
Cloud Credit Programs Federal credits for startups on open-model-optimized clouds (Azure, AWS, GCP, specialty) Startup access
Federal HPC Access DOE leadership computing (Frontier, Aurora, El Capitan) allocation for open frontier training Frontier pluralism
Semiconductor Incentives CHIPS Act R&D funding for open model optimization (quantization, sparsity, compilation) Cost reduction

2. Procurement & Standards (GSA, NIST, OMB, Agency CIOs)

Lever Policy Action Target
Federal Procurement Preference Mandate open-weight evaluation for AI contracts; require weight access for high-risk/sovereign use Sovereignty, competition
NIST AI Risk Management Framework Open-weight specific profiles (model cards, eval transparency, supply chain) Safety, trust
Model Card / Data Card Standards Mandate for federal use; align with HF/LF standards Transparency
SBOM for AI Software Bill of Materials extended to models (weights, training data provenance, dependencies) Supply chain security

3. Export Controls & International (BIS, State, NSC, USTR)

Lever Policy Action Target
Weight ≠ Software Clarification BIS guidance: model weights not "software" under EAR; different control list Avoid over-restriction
Allied Deployment Licenses General licenses for NATO/Five Eyes/major allies to deploy US-origin open weights Allied sovereignty
Deemed Export Rules Clarify foreign national access to open weights in US universities/labs ≠ deemed export Talent retention
Open Model Dialogue G7, OECD, bilateral: harmonize open model treatment; avoid fragmentation Global diffusion

4. Liability & Legal (Congress, Courts, FTC, State AGs)

Lever Policy Action Target
Safe Harbor for Release Good-faith pre-release eval + model card + responsible disclosure = liability shield Encourage responsible open release
Downstream Liability Allocation Clarify: base model provider not liable for fine-tuned derivative misuse (like Section 230 for models) Innovation
Distillation Protection Targeted civil/criminal remedies for unauthorized API extraction (not for legitimate distillation) IP protection without chilling
Open License Enforcement FTC guidance: restrictive licenses on "open" weights may be unfair/deceptive License honesty

5. R&D Funding (NSF, DOE, DARPA, ARPA-H, IARPA)

Program Area Open Weight Focus
Foundation Model Training Public compute for open frontier training (like DOE INCITE for supercomputing)
Efficiency Research Quantization, distillation, sparsity, compilation for open models
Evaluation Science Open eval frameworks, benchmarks, red-team methodologies
Domain Adaptation Open models for science (DOE), health (ARPA-H), defense (DARPA/IARPA)
Governance Research Technical alignment for open models; watermarking, provenance, unlearning

Licensing Policy Matrix

License Type Examples Commercial Use Derivative Distribution Patent Grant Policy Preference
Apache 2.0 Mistral 7B, Mixtral, OLMo, BLOOM ✅ ✅ ✅ Highest — OSI, patent grant, clear
MIT Some small models, component code ✅ ✅ ❌ (implicit) High — Simple, permissive
Llama Community Llama 2/3/3.1 ✅ (<700M MAU) ✅ (with attribution) ❌ Medium-High — Practical for most; 700M cap
Custom Restrictive Some Chinese models (Qwen old), proprietary "open" ⚠️ Limited ⚠️ Limited ❌ Low — Not truly open; procurement should prefer above

Policy recommendation: Federal procurement and funding should require or strongly prefer Apache 2.0 / MIT / Llama Community equivalents. Custom restrictive licenses should not count as "open" for policy purposes.

Release Norms (Industry Self-Regulation + Policy Nudge)

Model Tier Compute (FLOPs) Recommended Norms Policy Role
Small (<10^23) < 10B params typically Open by default; minimal gating Encourage via procurement
Medium (10^23-10^25) 10B-70B Staged release (red team → partners → public); model card mandatory NIST profile; voluntary
Frontier (>10^25) >70B, >10^25 FLOPs Coordinated disclosure; govt notification (voluntary); extensive evals; phased release Mandatory pre-release eval; BIS notification if dual-use

Key principle: Thresholds should be compute-based, not weight-availability-based. An open 405B model and closed 405B model pose similar capabilities risks; both should face same eval requirements.

International Alignment Challenges

Jurisdiction Approach Tension with US Open Weights Policy
EU AI Act GPAI obligations (eval, docs, copyright) apply regardless of openness; open weights get some carveouts EU may impose stricter openness definitions; copyright compliance harder for open weights
UK Pro-innovation; foundation model taskforce; favors open for competition Aligned; potential partner for standards
China Algorithm registration; security review for public-facing; open weights allowed but controlled Chinese open models (Qwen, DeepSeek, GLM) compete globally; US export controls may restrict US→China but not China→US
Canada AIDA (pending); voluntary code for generative AI Aligned; compute access via CIFAR/Digital Research Alliance
Japan Soft law; guidelines; sovereign AI push (Fugaku-LLM) Aligned; partner for open model development
Global South Limited compute; need open models for localization US open weights = digital public good opportunity

Measurement & Accountability

Indicator Data Source Frequency Target
Open model downloads Hugging Face, ModelScope, direct Monthly 2x YoY
Startup open model adoption PitchBook, YC, a16z portfolio surveys Quarterly >70% of AI startups use open base
Federal open model contracts USAspending.gov, SAM.gov Annual >50% of AI model contracts
Vulnerabilities found/fixed (open vs closed) CVE, GHSA, vendor advisories Annual Open ≥ closed discovery rate
Allied sovereign deployments State/DoD reporting, allied procurement Annual 20+ allied nations
Open frontier model count Epoch AI, Papers With Code Annual ≥3 orgs releasing >10^25 FLOP open models

Related Wiki Pages

Version History

Version Date Changes
1 2026-07-24 Initial policy framework from coalition statement