haskell
client<llm> Qwen3VL8B {
  provider openai-generic
  options {
    base_url "https://api.parasail.io/v1"
    api_key env.PARASAIL_API_KEY
    model "Qwen/Qwen3-VL-8B-Instruct"
    temperature 0
  }
}

function QwenChat(video: video?) -> string {
  client Qwen3VL8B
  prompt #"
    {{ _.role('user') }}
    Watch my video:
    {{ video }}

    Explain in detail what you see in the video
  "#
}

Unspecified error code: 2 Failed to create request builder: Error {
    context: "Failed to build request",
    source: "Video input is only supported on OpenAI's Realtime API (/v1/realtime), not on chat completions. Consider extracting frames from the video as images instead. See: https://platform.openai.com/docs/guides/realtime",
}

Request options: {"model":"Qwen/Qwen3-VL-8B-Instruct","temperature":0}

Qwen3-VL-*-Instruct

is amazing btw!! I was using GPT5 before for vision tasks, each prompt took freaking 30 seconds+ Qwen does it in 5 seconds

Thinking tokens are so annoying for agent tasks. The main limiting factor seems to be parameter size according to my anecdata, thinking mode just seems to slow down the generation with only marginal gains

It's amazing how in BAML I can just switch from GPT5 to Qwen3 with no issues. BAML just works man. If you take the time to craft your system and your promtps, BAML is so reliable

Let me take a look at how parasail accepts it