AI startups today are building on an unprecedented mix of foundation models — from open-source options like DeepSeek-R1, Meta's Llama 3, and Mistral's Mixtral, to proprietary systems like OpenAI's GPT, Anthropic's Claude, and Google's Gemini models.

The diversity and increasingly robust capabilities of open-source models have reduced the cost of product development and inference, while simultaneously introducing new, concrete IP risks that founders must proactively manage.

Why this matters now

1. Open-source models accelerate development — and increase IP complexity

Open-source systems give startups speed and cost-efficiency. Adoption numbers show this clearly:

  • DeepSeek-R1, released in January 2025, has gained attention for its low training costs and advanced capabilities compared to leading proprietary models.
  • Meta reports that the Llama family, including Llama 3, has seen widespread adoption, achieving over 1 billion downloads by March 2025, demonstrating the viral propagation of open-source AI code.

But legal researchers consistently warn that open-source AI does not eliminate patent risk:

  • Open-source code can still contain or be used to implement patented techniques; downstream startups can inadvertently infringe, as legal experts warn.
  • Widespread use of similar models increases parallel/overlapping inventions, which courts may treat as prior art affecting patentability.

2. Proprietary models shift the IP issues — but don't remove them

Using closed models like GPT-4o/5 or Claude removes some architecture-level risks — but introduces others:

  • API licenses may restrict derivative works or impose ownership limitations (e.g., limits on using outputs for competing models).
  • Because startups do not own or control the base model, reliance on proprietary systems can complicate investor diligence ("What exactly do you own?").

3. Investors are now heavily weighting IP maturity

Multiple venture analyses show IP as a key predictor of company success, requiring due diligence.

VC funding is increasingly tied to demonstrable IP defensibility.

  • Investor research indicates that strong IP positions reduce perceived commercial and regulatory risk, directly influencing valuation and deal terms.

Freedom-to-operate is now a gating issue.

  • With AI startups increasingly building on open-source models (e.g., Llama, DeepSeek, Mistral), FTO analyses have become a critical early gate for AI startups using open-source models. Without a clear FTO path, many deals stall in diligence.

Actionable steps for AI startups

1. Run an IP + foundation-model usage audit

Document:

  • What parts of your product use open-source weights
  • What is original (training methods, post-training pipelines, guardrails, inference engineering)
  • The licenses governing each component

Patent laws treat open-source disclosures as relevant prior art, directly affecting novelty.

2. File patents on specific technical improvements

The strongest patent claims today focus on:

  • Novel training techniques
  • Inference optimization
  • Architecture modifications
  • Alignment and safety innovations
  • Domain-specific improvements

This aligns with current USPTO guidance encouraging patents on technical AI improvements and practical applications.

3. Plan for truly global IP risk

Because open-source models propagate globally, ensure your strategy covers:

  • Patentability in the EU, UK, China (e.g., EPO's 2025 AI guidelines, which prioritize technical effects)
  • How disclosures abroad affect U.S. filings
  • Whether defensive publication is more appropriate for certain components

4. Integrate IP into fundraising and GTM planning

Well-structured IP improves:

  • Enterprise partnership readiness
  • Licensing flexibility
  • M&A valuation
  • Investor confidence

Reports from private capital market analyses note patent strength as a valuation multiplier for technology startups.

Bottom line

The rise of open-source foundation models — DeepSeek, Llama 3, Mixtral, and others — has made it easier than ever for startups to build sophisticated AI products.

But it has also increased the legal complexity around ownership, inventorship, and overlapping innovation.

Startups that take the time to document their originality, protect technical advancements early, and map global IP risks will be far better positioned to compete, differentiate, and attract investment — no matter which AI model or framework they are building on.