Hive Vision Language Model (VLM) brings flexible, prompt-based visual understanding to production vision tasks. With Hive VLM, teams can use natural language to define their policy and classify images, video, and text through a single endpoint.

But production workflows often need more than flexible prompting — they also need control.

That is why we are introducing Bias Tuning for Hive VLM, a new feature that allows teams to weigh certain classes more than others. By increasing or decreasing a class weight, teams can make that class more or less likely to be returned, helping them better align VLM behavior with their precision and recall goals.

Why Bias Tuning Further Sets Hive VLM Apart

Hive VLM is already built for best-in-class performance across image classification and moderation tasks, combining strong visual understanding with the speed and flexibility required for production workflows. Bias Tuning builds on that foundation by giving teams more control over how likely specific classes are to be returned.

This is a capability unique to Hive VLM, bringing built-in threshold refinement to enterprise-scale classification in a way standard VLMs do not.

Instead of relying only on prompt changes, teams can now tune class behavior directly inside the prompt. This is especially useful for workflows where the right operating point varies by policy, customer, or use case. A marketplace may want to reduce false positives when reviewing product listings. A social app may want to increase recall for harmful or inappropriate content. An advertising platform may want stricter controls for brand safety categories.

Bias Tuning makes those tradeoffs configurable in the prompt, with no retraining required.

How it works

The first step of using Hive’s VLM is drafting your prompt. Your prompt is your policy: changing it changes the output. A good prompt is explicit, defines important terms, covers edge cases, and gives the model the context it needs. Treat it like instructions for a human moderator: clear, consistent, and free of contradictions. For help formatting prompts, you can use the Prompt Builder: https://docs.thehive.ai/docs/prompt-builder.

With Bias Tuning, you can add a “biases” object anywhere in the prompt to weigh certain classes more than others. Increasing a class weight makes that class more likely to be returned. This increases recall and lowers precision for that class, helping teams tune the model to the right operating point for their workflow.

Once your prompt is ready, you add it to your API request. The Bias Tuning object lives inside the prompt itself, so it is included as part of the text prompt in the messages array along with the image or video you want the model to evaluate. The API runs the VLM with your instructions and returns the model’s answer plus token usage for the request.

This same flow also applies across text, images, and video. While text is processed directly, images are divided into patches so the model can understand objects, text, and spatial relationships. Videos are sampled into frames and analyzed either once or frame by frame depending on your sampling strategy and prompt scope.

Built for production content understanding

Teams can use Hive VLM for:

• Moderation: Identify harmful, unsafe, NSFW, or policy-violating content
• Tagging: Extract useful descriptors for search, ranking, and organization
• Detection: Find specific objects, logos, environments, behaviors, or visual signals
• OCR and text analysis: Read text in images and evaluate it within visual context
• Structured classification: Return outputs in JSON or other formats that fit production workflows

Try it yourself 

The best way to understand Hive VLM is to put it to work. Explore real examples in the Playground, try your own images, and see how easily you can customize prompts to match your needs. We are excited to see what you build with it. 

To help teams test Bias Tuning more easily, we are adding a preset example in the Hive VLM Playground and a quick-copy tooltip in AutoML Evaluations, giving them a simple way to see how class weights can be adjusted directly in the prompt.

If you want a version tailored to your brand or a deeper walkthrough, our team is here to help. Please reach out to our sales team (sales@thehive.ai) or contact us here for further questions.