Dream-1 robot API, distilled byte models, Ampere BF16 module

A thin day for lab launches, so this one is a robotics API, a paper about tokenizers and a small piece of systems work. I opened the primary page for each. I did not run any of them.

Dream-1 puts a robot policy behind an API

Dream Labs came out of stealth with $90M in seed funding and a World Action Model called Dream-1. The pitch is that you connect a bimanual robot with parallel grippers to their API and it starts working on tasks with no fine-tuning, no demonstrations from that robot and no machine learning background. The client code on the page is a handful of lines: pick the checkpoint, send an observation, get back an action.

The numbers on the page are the lab's own. Each call returns 1.6 seconds of actions, so an hour is about 2,250 calls. Dream-1 costs $3 per hour and the lighter Dream-1-flash costs $2. The team shows simulation results on 27 unseen robots and real clips from 12 partner sites, including the original PR2, a research robot released 18 years ago.

There is also an outside number. Robocurve, a public benefit corporation that measures robotics capabilities, ran an independent report on a Tier 1 benchmark. Dream Labs says Dream-1 scores above the next best models at 16x lower cost per task attempt than one of them. I have only the lab's summary of that report, so I would read the report itself before leaning on it. The page also says where the model still struggles and offers a fine-tuning service for tasks it misses, which is the honest part of the pitch.

Byte models that eventually beat token models

The paper from Marathe, Pagnoni, Limisiewicz, Li, Lewis, Zettlemoyer, and Iyer asks a simple question: if you distil a small model from a larger one, does it matter whether the small one reads tokens or raw bytes? The catch is that the teacher thinks in tokens, so the authors introduce two ways to convert token logits into byte logits, one approximate and one exact.

They then train models of roughly 1B parameters up to 1 trillion bytes of data. Token models win at low compute and then plateau. Byte models start worse and pass them with more compute. The distilled byte model matches the distilled token model using one-sixth of the training data, and storing teacher logits over 256 bytes costs about one-fifth as much as over a vocabulary of around 100K tokens. The scaling-law extrapolation is where the paper gets bold, predicting gains of up to 6.5 percent over Llama 3.2-1B. That is a forecast, not a measured result.

A kernel module for RTX 3000 cards

The last item is a Linux kernel module that changes the FP and BF16 instruction issue rate on GA102, GA103, GA104 and GA106 cards. It runs before the stock NVIDIA driver binds, so it does not need patching with each driver release. On a 3090 Ti the author reports an 8192 cubed GEMM going from 87.50 to 157.41 TFLOPS in BF16 and from 42.58 to 82.61 in TF32.

User reports in the repo put RTX 3090 gains between 25 and 67 percent, and note that decode, which is bound by memory bandwidth, barely moves while prefill, diffusion and FlashAttention improve. It changes how a card behaves at boot, so I would treat it as an experiment on a spare machine and read the compatibility table first.

What I'd watch next

Whether anyone outside Dream Labs reproduces zero-shot control on a robot that was not on the demo list. Whether byte distillation shows up in a model larger than 1B. Whether the BF16 module gets compatibility reports for more cards.