Skip to main content
Model APIs support text and vision inputs, but multimodal capability depends on the underlying model. Vision-capable models accept images alongside text in the same request, using the OpenAI-compatible image_url content type. The model processes both modalities together, so it can answer questions about image content, compare multiple images, or extract structured data from screenshots. Not all models support vision. Check the table below before sending image inputs.

Supported models

Nemotron 3.5 Super VL accepts text, image, and video inputs. Audio inputs aren’t supported through its Model API.

Send a vision request

Use the image_url content type to include images in your messages. Baseten retrieves image URLs from the inference service, so the URL must be reachable over HTTPS from Baseten’s environment. Prefer presigned object-storage URLs; network policies may block other hosts. Base64 data URLs are also supported and count toward the JSON request-body limit. Optional image_url.detail controls preprocessing resolution: low, high, original, or auto (OpenAI-compatible). When in doubt, use auto. Send an image alongside a text prompt like this:
vision.py

Image and video limits

These limits apply to images, videos, and audio sent to Model APIs. A model can impose lower limits or support fewer media types; check the supported models and Audio guides for model-specific requirements.
Selected clusters may temporarily accept JSON payloads larger than 16 MB, up to 100 MB. This allowance isn’t guaranteed for all models. Keep production requests at or below 16 MB.

Media counts by model

Kimi K3.1 Pro accepts a video_url content part, but video requests fail with an HTTP 500 error. Don’t send video to this model. Muse Spark Open accepts JPEG, PNG, WebP, GIF, BMP, and TIFF images. For animated GIFs, it uses only the first frame. It accepts MP4, MOV, MKV, and WebM videos with H.264, H.265, VP9, or AV1 encoding, subject to the video input limits. It samples at most 96 frames per video and ignores the audio track. Send speech as a separate audio_url part. A request can fit within the file-size limits and still exceed the model’s input-token or preprocessing budget. Split large combinations of images, videos, and audio across requests. Muse Spark Open doesn’t accept PDF files. Extract their text before sending it.

Send media by URL or base64

Send media as a base64 data URL or a direct HTTPS URL. Base64 encoding increases file size by about a third, and the encoded string counts toward the JSON request-body limit. Use URLs to keep larger files out of the JSON payload; the per-item and combined media limits still apply. For private media, use a presigned or signed URL from Amazon S3, Google Cloud Storage, Azure Blob Storage, Cloudflare R2, Oracle Cloud Object Storage, OVHcloud Object Storage, Vultr Object Storage, or IBM Cloud Object Storage. Keep the URL valid until the download completes. Other media hosts aren’t recommended because network policies may deny access, even when you can open the URL locally. Use a direct URL to the file and ensure the storage service can deliver it within the 5-second download limit.

Prepare videos before sending them

Use 1080p video at 30 frames per second or less. Source resolution and frame-rate limits apply before the model samples frames, so a small file can still be rejected if decoding it requires too much work. Many models process video at about 2 sampled frames per second, at resolutions up to 1080p. Sending a 4K video at 120 frames per second increases bandwidth and decoding work without necessarily providing more detail to the model. Resize and reduce the frame rate before sending the video. Contact us to discuss higher media limits for a dedicated deployment.

Pricing

There is no additional per-image fee. Images are converted to input tokens and priced at the model’s standard input rate. Higher resolution images produce more tokens and cost more to process. The exact conversion from pixels to tokens depends on the model. Kimi K3 divides each image into 28×28 pixel tiles where each tile becomes one input token. The cost table below uses Kimi K3’s uncached input rate ($3.00 per million tokens); for other models, use the rates on the Model APIs pricing page. For videos, token count scales with both resolution and the number of sampled frames.

Next steps

Model APIs overview

Supported models, pricing, and the feature support matrix

Chat Completions reference

Full request and response schema for the image_url content type