DEV Community

Cover image for NVIDIA Just Dropped an Open-Weight Speech Model. I Caught It on My Retro TV 1 Hour Later.
Shouvik Palit
Shouvik Palit

Posted on

NVIDIA Just Dropped an Open-Weight Speech Model. I Caught It on My Retro TV 1 Hour Later.

I wasn't looking for it. I was just having my morning coffee with DEV·TV running on the side.

Then channel 11 flipped to a new bulletin:

NVIDIA: Nemotron 3.5 Transcribe. Added 1 hour ago.

![DEV·TV showing Nemotron 3.5 Transcribe on the NEW MODELS channel]
NVIDIA

What NVIDIA released

Nemotron 3.5 Transcribe is NVIDIA's new speech recognition model, and the specs are worth a look:

  • 🔓 Open-weight. You can download it and run it yourself.
  • 🧠 30B mixture-of-experts, with only ~4B active parameters. That gives you the knowledge of a big model at roughly the running cost of a small one.
  • ⏱️ Transcribes recordings up to an hour long. That covers full meetings, podcasts and lectures in one pass.
  • 🌍 Supports 40 languages.

Why this one matters

Speech-to-text has mostly meant two choices: pay for a closed API, or settle for a smaller open model. An open-weight, mixture-of-experts transcriber from NVIDIA changes that.

With only ~4B parameters active at a time, it should be far cheaper to run than its 30B size suggests. If you're building meeting notes, podcast search, captioning or voice agents, put this on your shortlist.

And it wasn't alone

Right after it, DEV·TV queued up two more audio-related releases:

  • NEXT: AssemblyAI TTS 1
  • LATER: Particle AI Particle-1

Three fresh models, back to back, without me refreshing a single page.

How I caught it so fast

I built DEV·TV, a retro TV for developers that runs in your browser. Its NEW MODELS channel pulls fresh listings from OpenRouter and airs them like a live news bulletin.

Most model launches happen quietly. A listing appears, a few people notice, and by the time it trends the first wave is over. With DEV·TV on my second monitor, the new models come to me.

It also has:

  • 🥊 MODEL FACE-OFF: New models compared head to head
  • 🛠️ AGENT SKILLS: Trending Claude skill repos from GitHub
  • 🌟 SPOTLIGHT: Indie dev projects, free to submit from the app

The entire app is one HTML file with zero infrastructure, and it's open source. It's been #5 Product of the Day on Product Hunt and #1 on IndieNeed.

Tune in

Want to catch the next big model drop the hour it lands?

👉 shouvik12.github.io/devtv

If you try Nemotron 3.5 Transcribe, tell me in the comments how it handles your audio. 🎙️


Animated cover GIF made with living-canvas. Turn any repo or idea into an animated banner.

Top comments (0)