Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on @huggingface, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

Joe Nemotron posting in all caps today. I have spent a *lot* of time tinkering with the new Nemotron 3.5 Super VL and can confirm that it's a meaningful step up in performance for the 120B (12B active) category. As always with Nemotron models, it's a great base for custom post-training/fine-tuning because NVIDIA releases the training recipes and most of the model's training data as open source. This is very helpful in iterating on, for example, synthetic data generation during fine-tuning. (A typical question during fine-tuning is "do I want to generate a bunch of data for this particular metric or mix in data the model saw in pre-training?") This model also has vision capabilities. I'm looking forward to seeing @skalskip92's analysis of how the "VL" in Nemotron Super 3.5 VL compares to other recent models.

Going live in 5 minutes!

More multi-model voice AI demos from Jon: this time a tiny bi-directional encoder model in the Pipecat pipeline that runs on CPU fast enough to be in the main voice loop. Fun fact: you can train a classifier model like this in less than 10 minutes on a modern GPU. (If you have the right data set. All ML problems are data problems.)

I really like the Bland Speech v3 voice model from Bland, and a lot of other people do too! It's currently the top realtime TTS model on the Design Arena Audio Realism ranking, which is a tournament-style benchmark in which actual people judge which model sounds more human. The model is also quite fast. In my evals, I'm seeing ~200ms P50 time-to-first-speech and 400ms P95. That's top 5 among the ~20 text-to-speech models I monitor.

Add to the list of things I did not expect to be true when computers became super-intelligent: the single most important glue-layer tool (load-bearing, as the LLMs say) would be a terminal multiplexer written in 2007.