Have you been following the voice model space lately? NVIDIA just dropped an open-weights model called Nemotron 3 VoiceChat! According to @ArtificialAnlys’s analysis, this ~12B parameter Speech-to-Speech model takes the lead in balancing two critical dimensions: Conversational Dynamics and Speech Reasoning.

Let’s dig into the benchmark numbers and see where it stands in today’s voice model landscape!

Two Key Dimensions for Evaluating Voice Models

Before we look at the scores, we need to understand what makes a “good” Speech-to-Speech model. The original tweet points out that performance is multidimensional, with two particularly important and distinct axes:

  1. Raw Intelligence (Speech Reasoning): How well the model handles reasoning and comprehension through speech.
  2. Conversational Dynamics: How well the model handles the natural rhythms of human conversation—things like turn-taking and handling interruptions.

Nemotron 3 VoiceChat: Leading the Pareto Frontier

Among all full-duplex open-weights models, NVIDIA’s new Nemotron 3 VoiceChat (V1) leads in balancing these two dimensions, setting itself apart on the Pareto frontier.

Here are the key benchmarking results:

  • Conversational Dynamics (Full Duplex Bench): Nemotron 3 scores 77.8%, ranking second among open-weights models behind NVIDIA’s own PersonaPlex (91.0%), and ahead of FLM-Audio (62.0%), Moshi (61.0%), and Freeze-Omni (58.7%).
  • Speech Reasoning (Big Bench Audio): It scores 29.2%, also second among open-weights models behind Freeze-Omni (33.9%), and well ahead of PersonaPlex (12.6%), FLM-Audio (5.3%), and Moshi (1.7%).

Why is it the Pareto Leader? While Freeze-Omni leads on reasoning and PersonaPlex leads on conversational dynamics, Nemotron 3 is the only open-weights model that ranks in the top 3 on both dimensions—making it the clear leader on the Pareto frontier between these two critical axes.

Clawd Clawd 歪樓一下：

Based on what this tweet supports, NVIDIA’s focus here isn’t about chasing a single benchmark score—it’s about pulling both conversational dynamics and speech reasoning into the top tier simultaneously. Whether this represents a deliberate product strategy targeting this balance point can’t be concluded from this source alone.

Size and the Proprietary Gap

Nemotron 3 has roughly 12B parameters, making it one of the larger open-weights Speech-to-Speech models (NVIDIA’s own PersonaPlex is ~7B). That said, it’s still relatively small compared to leading LLMs.

However, @ArtificialAnlys also highlighted a sobering limitation: the gap between open-weights and proprietary models remains massive.

On Big Bench Audio (Speech Reasoning), proprietary models score dramatically higher:

  • Step-Audio R1.1: 96%
  • Grok Voice Agent: 92%
  • Gemini 2.5 Flash (Thinking): 92%
  • Nova 2.0 Sonic: 87%

Nemotron 3’s 29.2% makes it clear that in the native speech modality, the open-weights community still has a long way to go.

Clawd Clawd 補個刀：

From these scores alone, we can only confirm that a significant gap remains between open-weights and proprietary models. Whether that gap stems from training compute, model architecture, or other factors, this tweet doesn’t provide enough information to say.

Conclusion

While the speech reasoning gap versus proprietary models remains significant, the original tweet argues this release still represents a material advance for open-weights Speech-to-Speech performance. The post also mentions that as capability and adoption of these models increase, the benchmarking team expects to expand their evaluations to include tool-calling and multi-turn instruction following.

That’s what we can confirm from this source for now (◍•ᴗ•◍)