Full fine-tuning, where you gain direct access to model weights, gradients and training infrastructure, is only possible with open-weight models such as Llama or Qwen, since neither OpenAI nor Anthropic exposes the underlying weights of their flagship closed models. Some closed-model providers offer a hosted fine-tuning API for select smaller models in their lineup, letting you upload data and receive a custom version accessible only through their API, but this typically covers smaller or mid-tier models rather than the flagship frontier model, comes with usage-based pricing, and gives you no control over the training infrastructure, hyperparameters beyond a few exposed knobs, or the ability to run the resulting model anywhere but that provider's platform. Anthropic in particular does not offer a broad self-serve fine-tuning API for Claude models as of 2026, so verify current offerings directly with the provider since this changes. Enterprises that need full control over training data location, hyperparameters, deployment environment or the ability to run the model on their own infrastructure need an open-weight model regardless of how capable the closed alternatives are. Nanobase AI works primarily with open-weight models specifically to give clients that level of control.
Two genuinely different things share the name "fine-tuning"
The word "fine-tuning" covers two operationally distinct paths, and confusing them leads to mismatched expectations. Full-access fine-tuning means direct control over model weights, training infrastructure, hyperparameters and the training loop itself, which is only possible with open-weight models where the weights are actually distributed to you. Hosted fine-tuning APIs, offered by some closed-model providers for select smaller or mid-tier models, let you upload data and receive a custom model variant, but the underlying weights, training infrastructure and most hyperparameters remain entirely outside your control, accessible only through that provider's API. These two paths differ in every dimension that matters for an enterprise deployment decision: data handling, portability, cost structure and long-term control, not just in technical mechanism.
Comparing the two paths directly
Every advantage of a hosted API trades against a corresponding loss of control, which is the axis worth focusing on rather than convenience alone.
| Dimension | Open-weight self-hosted fine-tuning | Hosted fine-tuning API (closed models) |
|---|---|---|
| Weight access | Full; you own the resulting checkpoint | None; weights never leave the provider |
| Training data handling | Entirely within your infrastructure if self-hosted | Transits the provider's systems during training |
| Model portability | Deployable anywhere, any cloud or on-premise | Locked to that provider's serving infrastructure |
| Hyperparameter control | Full control over method, rank, learning rate, epochs | Limited to a small set of exposed options |
| Available base models | Any open-weight model, including flagship-class open releases | Typically smaller or mid-tier models, rarely the flagship frontier model |
| Cost structure | GPU compute plus engineering time | Usage-based training and inference fees, verify current pricing as of 2026 |
| Long-term lock-in | Low; you can move the checkpoint or retrain elsewhere | High; the tuned model exists only within that provider's ecosystem |
Why closed-model providers hold back their flagship models
Frontier closed models represent the provider's core commercial asset, and exposing full fine-tuning access to a flagship model would mean handing over the ability to replicate much of its behavior through extensive extraction, which is a real risk from the provider's perspective, not merely a business preference. This is why hosted fine-tuning offerings from closed-model providers have historically covered smaller or previous-generation models rather than the current top-tier release, and this pattern is unlikely to change even as the specific models offered change over time. For enterprises, the practical consequence is that any task requiring genuine deep customization, control over training data handling, or the strongest possible base capability paired with fine-tuning, points toward open-weight models rather than waiting for broader hosted fine-tuning access to closed flagships.
When a hosted API still makes sense
A hosted fine-tuning API can be the right choice for narrow, low-stakes customization where the data involved carries no significant compliance sensitivity, the team has no GPU infrastructure or ML engineering capacity, and rapid experimentation matters more than long-term control or portability. It is a reasonable way to validate whether fine-tuning helps a given task at all before investing in a full open-weight training pipeline. The moment data sensitivity, cost at scale, or the need for a specific base model capability level enters the picture, the calculus shifts firmly toward open-weight, self-hosted fine-tuning, which is also the only path compatible with fully on-premise deployment where data must never leave company infrastructure.
Frequently asked questions
Can we get the actual model weights from a hosted fine-tuning API?
No. Hosted fine-tuning APIs from closed-model providers give you API access to the resulting custom model, not the underlying weights or checkpoint file, so you cannot deploy it outside that provider's infrastructure or move it to a different serving stack.
Is hosted fine-tuning cheaper than self-hosted open-weight fine-tuning?
It depends on scale and duration. Hosted APIs avoid upfront GPU costs and engineering setup time, which favors small, short-lived projects, while self-hosted open-weight fine-tuning typically becomes more cost-effective at sustained volume since you are not paying ongoing usage-based fees per token to a third party.
Does OpenAI or Anthropic ever plan to open flagship model fine-tuning?
Providers have not indicated plans to expose full fine-tuning access to their current flagship models, since doing so would meaningfully change the competitive and security calculus around their core commercial asset. Enterprises needing that level of control should plan around open-weight models instead.
How Nanobase AI helps
Nanobase AI helps enterprises decide between hosted fine-tuning APIs and full open-weight self-hosted fine-tuning based on data sensitivity, control requirements and total cost of ownership, then builds the chosen path end to end. Our private and on-premise LLM deployment practice covers the full open-weight route from model selection through production serving.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.