Closed APIs, downloadable model weights and shared protocols offer different kinds of access. From GPT-3 and LLaMA through DeepSeek and later Chinese open-weight releases, this Arc examines what changed in availability and licensing—and what remains a company-reported performance claim. A market selloff after a model release does not prove single-cause attribution; published weights do not automatically grant unrestricted rights or show widespread use. The question is who can run, modify and distribute a model, under which conditions.
GPT-3 was offered through an API rather than a public download of its model weights. DALL·E 2, GitHub Copilot and ChatGPT likewise illustrate ways people could use a product without receiving the underlying model. These arrangements differ in pricing, interface and intended use; calling them all “closed” can obscure those differences. They also do not establish that no downloadable models existed at the time, or that compute alone determined each release decision. The useful comparison for this Arc is narrower: service access can be valuable without granting permission to inspect, run independently or redistribute model weights. Later releases tested different combinations of those permissions.
Key Insight
Access to a model-backed service is not the same as access to its weights or a license to redistribute them.
Stable Diffusion’s release in 2022 gave people another way to run image generation outside a single hosted service. Meta made the original LLaMA weights available to approved researchers under a restricted license in 2023; unauthorized redistribution was not the same thing as a permissive open-source release. Later Llama models, including Llama 3.1 405B, broadened the range of downloadable weights, though a model’s license and deployment costs still determine what users can do. These events show multiple ways AI systems can be shared; none alone establishes a universal capability ranking or a single cause for subsequent research.
Key Insight
Research access, unauthorized copying, licensed weight downloads and application code involve different permissions.
DeepSeek released R1 in January 2025 with downloadable weights and published technical claims about reasoning performance and training efficiency. Financial reporting also described a sharp Nvidia share-price fall days later and cited DeepSeek among investors’ concerns. A market move has multiple potential drivers; timing and commentary cannot establish how much of the price change the release caused. Nor does a company’s stated training cost define the full cost of comparable systems under equal accounting. Llama 4 and GPT-5 are additional model releases in this chapter, not controlled benchmark subjects for a direct R1 comparison. R1 changed what users could inspect and run; whether it matched any particular proprietary model depends on independently specified tasks, measurements and deployment conditions.
Key Insight
A downloadable release and a large contemporaneous selloff are observable; cost superiority and single-cause market attribution require separate proof.
A company-specific US export directive led Anthropic to suspend access to Fable 5 and Mythos 5 in June 2026; the restrictions were lifted June 30, and Fable 5 first-party access resumed July 1. After an earlier initial rollout, Z.ai made GLM-5.2’s MIT-licensed weights available on June 16; DeepSeek V4 had appeared earlier in April. These are different decisions by different actors; their sequence cannot show that the directive caused either release. GLM-5.2 and DeepSeek V4 illustrate alternative ways to make weights available, but vendor-reported benchmark and cost figures are not independent proof that they matched a proprietary frontier model. Licensing and technical access vary between releases, and geographic access rules may differ from product policies. The existence of an alternative is relevant to discussions of access, not evidence that export controls were ineffective or that unrestricted cross-border adoption followed.
Key Insight
The proximity of restrictions and weight releases raises an access question, not evidence of a causal policy reversal.
The distinction between model access and model openness became harder to reduce to a binary choice. OpenAI’s public-benefit-corporation restructuring, Sora’s discontinuation, the Model Context Protocol, and OpenClaw concern different layers of the industry: ownership, product continuity, interoperability, and software distribution. None of these events alone demonstrates that open and proprietary models reached performance parity. On October 15, StepFun made Step 5 Preview weights available under its Model License 1.0. The model repository describes self-hosted inference, while the license adds attribution and commercial-use conditions. It is an example of available weights with conditions, not unrestricted open-source licensing; neither its availability nor company-reported benchmarks establish independent performance parity or broad production adoption. The open–closed question therefore also concerns who can deploy, modify, and redistribute a model, and under what terms.
Key Insight
Available weights do not by themselves establish unrestricted redistribution, independently verified performance or production adoption.
Conclusion
The examples in this Arc distinguish access to a service from access to weights, and both from permission to reuse them. LLaMA’s limited research release, later downloadable models and Step 5 Preview’s conditional license did not confer the same rights. DeepSeek R1 was discussed alongside a large market selloff, but a temporal sequence cannot isolate its contribution to price changes. Likewise, selected or company-run benchmarks cannot establish general parity between model families. Open and proprietary approaches may coexist; neither licensing labels nor a single release can settle which approach will prevail. Future comparisons need independent testing of defined tasks, evidence of actual adoption and close reading of the relevant terms.
