If you've ever wanted to take a pre-trained language model and «tweak» it for your specific needs – for instance, teaching it to respond in a certain format, understand images, or use external tools – then you're already familiar with the idea of fine-tuning. It's a process where a model isn't trained from scratch but is instead «tailored» to a specific domain or use case. This is exactly the area Together AI is developing, and the company recently expanded its fine-tuning service significantly.
What Is Fine-Tuning and Why Is It Needed?
Simply put, fine-tuning is when you take a large, already capable model and further «train» it on your own data. As a result, the model becomes better at handling your specific tasks while retaining all its previous capabilities.
This is useful in many different situations: companies want a model to respond in their specific corporate style, developers aim to teach it to work with particular systems and APIs, and researchers seek to elicit specific behaviors in a narrow domain.
Together AI offers this service in the cloud: no need to set up your own infrastructure, rent servers, or get bogged down in configuration details. You upload your data – you get a fine-tuned model.
Three New Areas: Tools, Reasoning, and Images
The core of the update is three new types of support that were previously missing from the service.
Tool Calling. Modern language models can do more than just respond with text; they can also «call» external functions, like searching a database, sending a request to an external service, or performing a calculation. This is known as tool calling. Previously, fine-tuning a model for this specific behavior on the Together AI service was difficult. Not anymore. Now, you can train a model to correctly formulate these calls, process responses, and build scenarios where it interacts with the external world.
Reasoning. This refers to a category of models that «think out loud» before giving an answer – they go through intermediate steps of reasoning. This approach improves the quality of responses to complex logical, mathematical, or multi-step problems. Together AI now supports fine-tuning for this behavior, meaning you can train a model not just to provide a result, but to show its train of thought.
Vision-Language Models. These are models capable of processing not just text, but images as well. For example, they can take an image and a text-based question as input and return a meaningful answer. Fine-tuning for these models is now also supported, opening up opportunities to create specialized systems, from medical diagnostics using scans to analyzing documents with charts and diagrams.
Larger Models and Higher Speed
Beyond new task types, the update also addressed the technical side.
First, the service now supports fine-tuning for models with over 100 billion parameters. If that number doesn't mean anything to you, imagine this: the more parameters a model has, the more powerful and versatile it typically is, but also more demanding on resources. Working with such models was previously accessible only to large teams with serious infrastructure. Now, it's available via a cloud service.
Second, throughput has increased by up to 6x. In simple terms, training has become faster – data is processed more efficiently, reducing the waiting time for your results.
Knowing How Much It Costs and When It Will Be Done
One of the new practical features is the ability to estimate the cost and completion time of a job before you run it. This might sound like a minor detail, but it's crucial in practice. Previously, developers would launch a fine-tuning task without knowing exactly how long it would take or how much it would cost. Now, you can see the estimated figures before starting and make an informed decision.
This is particularly useful for teams working with tight budgets or deadlines. Uncertainty in timing and costs is a major source of friction when using cloud ML services, and removing it makes the entire process significantly smoother.
Who Does This Really Change Things For?
Looking at the big picture, the Together AI update makes fine-tuning more accessible to a broader range of teams. Before, working with powerful models, visual data, or complex tool-use scenarios required either substantial technical resources or complicated workarounds. Now, it's all packaged into a single service.
For small teams and independent developers, this means they can build quite sophisticated products without spending months setting up infrastructure. For more mature companies, it's an opportunity to iterate faster and test hypotheses without incurring extra costs.
Fine-tuning is gradually shifting from an expert-level procedure to a tool that can be used without a deep understanding of how the models themselves work. And that, perhaps, is the most important trend this update confirms.
