Dessa, Hy-Nomas, TensorRT 11.1 tuner, Microduck

A quieter day for frontier launches, but two models and two NVIDIA engineering posts are worth reading. I opened each primary page. I did not run any of the models.

Dessa, a model that learned to talk from sound alone

Block's small research lab says Dessa was trained from scratch on human conversation with no text and no pretrained models. It is a speech-to-speech model that is full-duplex, so it listens and talks at the same time, and it learns when to cut in or say "yeah" without being told.

The page makes several claims I cannot check: that inference runs faster than real time, that it runs locally on a MacBook Pro M4 Max or better, and that training fit on a single NVIDIA GB200 NVL72. Access is a waitlist, and a small group of safety researchers gets early access. The audio samples are the evidence for now, and the lab frames the work as an open research question about how efficiently speech alone teaches language.

Hy-Nomas, an open world model

Tencent released Hy-Nomas under Apache-2.0. It generates a video world one segment at a time, and a persistent key/value cache carries context between segments, so you can change objects, style or camera as it goes. The distilled configuration uses 8 denoising steps and outputs 720 by 1280 at 24 FPS. A separate configuration targets a single GeForce RTX 4090 using INT8 DiT weights, an INT8 KV cache and CPU offload.

Tencent also published Hy-Nomas-Bench, four benchmarks with 280 test cases covering multi-turn text interaction, camera navigation, mixed control and spatial memory. A benchmark released with the model is useful, though it is the authors' own.

TensorRT 11.1 gets a tuner

Part 2 of NVIDIA's series covers the Global Performance Tuner. A build route is a set of compiler knob values, and TensorRT 11.1 exposes more than 200 knobs. The trtexec tool expands a route expression, builds each candidate, measures it, and throws out any build that crashes or fails accuracy. The fastest eligible engine wins. Across ten workloads the gain over the default route ran from 1.14x to 1.63x. The Python version is slated for Torch-TensorRT 2.15.

I like that accuracy is a filter and speed is the ranking, not a blended score.

Microduck in simulation

The second NVIDIA post is a tutorial rather than a release. Agents on an RTX Spark plan the project and build assets, Isaac Sim assembles the obstacle course, and Isaac Lab trains a perception and locomotion policy across 4,096 parallel environments. Sim-to-real transfer is described as the next step, so no hardware result is claimed.

What I'd watch next

Whether Dessa ships beyond the waitlist, and whether anyone reproduces its speech-only training claim. Whether Hy-Nomas-Bench numbers hold up on other people's runs. Whether the TensorRT gains show up on my own models once the Python tuner lands.