AWS Guide: How to Use Slow Vision Encoder with Grok 4.3 on Bedrock Mantle

Select Language:

If you’re working with image or vision requests using the xai.grok-4.3 model through the Bedrock Mantle OpenAI-compatible endpoint, you might notice it takes quite a while to get a response. In fact, the time to first token can range from about 39 to 46 seconds for a single photo. In comparison, text-only requests usually respond in just 1 to 2 seconds, and similar requests with the moonshotai.kimi-k2.5 model on the same image in the same region typically take around 5 seconds.

To understand why this happens, it’s helpful to look at some tests. When sending a simple text prompt like “say hello,” the response is very quick—about 1 second. However, if you send a small 1×1 pixel PNG image (about 111 bytes) along with your prompt, the response time climbs to around 1.9 seconds. For a real JPEG image that’s 211 KB, the response can take as long as 39 seconds, despite the prompt size remaining the same. Even when the JPEG is cut down to 25 KB, the time only slightly increases to about 45.6 seconds. Interestingly, the number of tokens decoded from the image—around 1512 tokens—seems to be the main factor affecting latency, not just the image size on the wire or other processing steps.

This pattern suggests that the delay is linked to processing the image tokens, and the latency can vary between about 39 to 46 seconds. It also indicates that the vision backend might be experiencing slow or cold starts behind the Mantle layer, leading to these delays.

On the other hand, when you use the same 211 KB image with the Grok 4.3 model through the direct xAI API, the output is much faster, taking only 3 to 5 seconds.

If you’re looking for help, it’s good to ask about the typical response time expectations for vision requests with xai.grok-4.3. Also, check if there are any known issues related to slow starts or capacity problems in the Mantle version of the service, especially on the image processing path. Knowing this can help you plan your workflow better or reach out for support if needed.