Nemotron 3 Super opens up and Gemini Embedding 2 goes multimodal

Not much shouted for attention this week, which suits me. The quieter releases are often the ones I end up using.

Nemotron 3 Super

Nvidia released Nemotron 3 Super on 11 March. It is a 120B-parameter mixture-of-experts model that activates 12B parameters per pass, with a hybrid Mamba and Transformer design and a 1M-token context window. Nvidia is aiming it at agent workloads where many calls run in parallel and throughput decides the bill.

The part I find more interesting than the specs is how open it is. Nvidia released weights along with data, training recipe and infrastructure details. That level of disclosure is rare from a company of this size, and it gives researchers and enterprises something to reproduce and audit, and not only download it.

Why would Nvidia, whose business is selling chips, publish models? The plain answer is that good open models create demand for the hardware that runs them. I do not need a conspiracy for that, and the result is still useful to me.

Gemini Embedding 2

Google released Gemini Embedding 2 on 10 March. It turns text, images, video and audio into vectors in one shared space. Earlier setups needed a separate model for each kind of material and a messy layer to combine results.

Think about what that means for a project archive. A drawing, a photo of an installation, a voicemail and a PDF spec can all be searched with one query. I would test it on a slice of real files first, since embedding quality varies a lot by domain and I have seen strong demos fall apart on technical documents.

Smaller open tools

MatAnyone 2 cleans up subject extraction from video, and Fish Audio's S2 Pro offers an open voice model with controllable delivery. Both are the kind of practical tool content people will pick up fast.

Because it was a thin week I used the time to rerun my own checks on two of the January and February models, and that was more informative than the news. A couple of the models that looked best on launch day were mediocre on my documents, and one that got little attention did well. It is a reminder that public scores describe the average task, and my tasks are not average. They are scanned, bilingual and full of unit conversions.

What I'd watch next: third-party tests of Nemotron 3 Super on long agent tasks, and real pricing on Gemini Embedding 2 for video.