Nvidia open-sourced a dual-tower model on september 18 that writes 2.4 times faster than its…
nvidia open-sourced a dual-tower model on september 18 that writes 2.4 times faster than its single-tower baseline with only slight drops on code and math. the two towers generate text chunks in parallel and need two h100 or a10 gpus to run.
open weights just made speed, not score, the selling point.
Context
The NVIDIA model card for Nemotron-Labs-TwoTower-30B-A3B-Base-BF16 and the NVIDIA paper arXiv 2606.26493 describe a base block-wise autoregressive diffusion language model with two 30B towers. They report 2.42x the autoregressive baseline's wall-clock generation throughput at the default operating point on 2xH100 in BF16, with 98.7% of aggregate benchmark quality. The card's table, baseline to TwoTower, shows HumanEval 79.27 to 75.58, MBPP-Sanitized 74.71 to 74.28, GSM8K 92.49 to 90.14 and MATH-500 84.40 to 80.60. It lists 2x A100 80GB or 2x H100 80GB, about 59 GB per GPU in BF16.
All figures are NVIDIA's own on base-model benchmarks. 2.4x rounds 2.42x. The measured drops are 0.4 to 3.7 points, and the paper calls them modest; slight is the note's word. The card lists A100 or H100, and A10 is not in the text read. The card is governed by the NVIDIA Nemotron Open Model License Agreement, which was not read. No first-party text says 18 September: the card's model dates run September 2025 to April 2026, and the paper identifier and secondary reports of 1 and 2 July 2026 point to an earlier release. Speed, not score, as the selling point is the author's opinion.
Watch next
- Independent throughput reproductions and an instruct variant.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 19 September 2026 at 04:05 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →