Deep Principle published the BF16 weights of Hydrogen-1.0, a science model built on GLM-5.3. The rest of the week is research: a climate emulator, two robot papers and a voice model, none of which you can put to work on Monday.
Deep Principle releases Hydrogen-1.0, a science model built on GLM-5.3
Illustration generated with AI on our own machine.
What happened. Deep Principle built Hydrogen-1.0 on GLM-5.3, and the repository holds its unquantized BF16 weights (Hugging Face). In Deep Principle’s demos it explores a rearrangement network, generating 27 molecular structures through 30 plausible reactions, and writes a JSON protocol for a 12-vial Suzuki coupling study (Deep Principle).
What we think. We like that the demo ends in a lab protocol a chemist can check, not a chat answer. The model card says it was tested with SGLang on 8 GPUs (Hugging Face), so this is a laboratory’s server, not a laptop. The comparisons in the launch post are Deep Principle’s own.
The good bit. Open weights let a lab run it next to data it cannot send anywhere.
What you can do. If you are not a lab, nothing yet. If you are, our guide to how much VRAM a local model needs tells you what 8 GPUs means in practice.
Nvidia publishes STRATA, an AI emulator of a kilometre-scale climate model
Illustration generated with AI on our own machine.
What happened. Nvidia published STRATA, a global, kilometre-scale AI atmosphere model that emulates the SCREAM storm-resolving simulation (Hugging Face). It predicts the atmosphere on a grid of about 25.2 million columns with a 10-minute time step. It takes 149 input channels and produces 148 output channels.
What we think. We think emulators like this are where AI earns its keep in science: they copy an expensive simulation so it can be run many more times. A kilometre-scale grid sees storms that coarser models smear out. It is a tool for weather and climate centres.
The good bit. The model card lists every input and output channel, which is what a researcher needs to trust it.
What you can do. Nothing for a business this week. If you build models yourself, see our local AI for developers page.
DreamX learns robot actions from human and robot video
Illustration generated with AI on our own machine.
What happened. DreamX repurposes a pretrained video diffusion model as a universal latent action learner (Hugging Face). Its authors pretrained it on 2.19k hours of human and robot video, with about 300 times compression via learnable queries reading video features (DreamX). The paper reports the best results among latent action models on LARY-Bench.
What we think. We read “the best results” as the authors’ claim on a benchmark of their field, not a robot in your warehouse. Learning from human video matters because there is far more of it than robot footage.
The good bit. If robots can learn from ordinary video of people working, training them gets cheaper.
What you can do. Nothing to install yet. To judge claims like this on your own tasks, read how to benchmark a model on your own machine in 30 minutes.
ZAP matches GPT-6 Astra on robot tasks at lower cost, its authors say
Illustration generated with AI on our own machine.
What happened. The ZAP authors say that on real-world tasks from RoboLab-120 it matches GPT-6 Astra running in their harness, and that it executes each tool call much faster and cheaper (ZAP). In the same harness, the base Qwen 3.6 27B succeeded on 8.3% of RoboLab-120 scenes and 16.7% of real-world trials.
What we think. We find the base model’s score the most useful number here: 8.3% shows how hard these tasks are for an ordinary model. “Matches Astra” is the authors’ own measurement in their own harness. A cheaper tool call is what makes a robot affordable to run all day.
The good bit. The authors publish the weak baseline alongside the strong result.
What you can do. Nothing for an office yet. If you run models in-house, our local AI for developers page is the starting point.
MOSS-Live tops its own full-duplex voice benchmarks
Illustration generated with AI on our own machine.
What happened. The makers of the voice model MOSS-Live say it ranked first among the models they evaluated on six full-duplex metrics (MOSS). Its tool calling success rate on FDB v3 was 51%, tied for the highest.
What we think. We note these are the makers’ own tests. A 51% tool calling rate is a top score and still half of the calls, which tells you how young voice agents are.
The good bit. Full-duplex voice is what makes a phone assistant feel less like a menu.
What you can do. Wait before you put a voice agent on your customer line. When you do, check our local AI for developers page for keeping the audio on your own network.
In one line
Hydrogen-1.0 ships open BF16 weights for chemistry; we watch it as research, not as a tool for your office.
Work with us
If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.