Does anyone tried to train it on your Spark already? I’m comparing it right now with Opus 5, but I think it’s very optimistic claim that it’s going to beat it.
@ululusem, not trained here, only served, but if you are comparing it with Opus 5 on your own tasks, this is what one Spark gives you to compare: the full 1,048,576-token window (a code hidden at 10, 50 and 90 % of prompts up to 1,011,142 tokens was found 12 times out of 12), tool calls right 20 of 20 with the reasoning on and off, images and video, about 70 tok/s rewriting code with one request. We serve it as one lane of dgx-spark-qwen38, behind the same OpenAI- and Anthropic-compatible endpoint as the other models of the box, so you can run the same tasks against it: curl -fsSL https://raw.githubusercontent.com/hasso5703/dgx-spark-qwen38/main/get.sh | bash. Measurements and method: docs/nemotron-super-lane.md.
You use qwen38 in name, I assume it’s just name of the repo ::D Anyway, thank you. I still trying to get more time to test it properly Nemotron vs Opus and will report my test results when I finally do it.
@ululusem Yes, only the name: the repository started with Qwen3.8 on the Spark and kept the name as it grew; it now serves Nemotron 3.5 Super VL and Bolt too, behind the same endpoint. Looking forward to your results.