China's DeepSoma simulates whole brains inside real physical worlds
Whole-brain emulation took a step from viral stunt to would-be platform on Saturday, and a tiny on-device model quietly showed phones can run serious AI fully offline.
China's DeepSoma is building a whole-brain simulation platform that drops a complete biological brain into a real, changing world. The startup Zhiyue Space Intelligence unveiled the system as a direct challenge to Eon Systems, the San Francisco neurotech firm that grabbed headlines in March by loading an adult fruit fly's connectome — around 140,000 neurons and 50 million synaptic connections — into a virtual body and letting it forage and groom inside a physics simulator. Eon's demo was deliberately simplified: its fly runs on leaky integrate-and-fire neurons, part of the brain-to-body mapping is hand-tuned, and only a narrow slice of sensory input is covered.
DeepSoma aims to go three levels deeper. Its model layer rebuilds individual neurons biophysically — dendrites, soma, membrane potential, ion channels and synapses — instead of treating them as firing aggregates, then wires those cells into circuits and a whole connectome. Its world layer rejects preset physics scenes in favor of reconstructing real environments as continuously updated, computable 4D digital worlds that carry geometry, semantics and physical information. And its platform layer, which the company sums up as "Build Worlds, Run Brains, Embody Intelligence," lets the same brain model drive a digital animal, a lab experiment, or a robotic arm and humanoid, closing the loop from the environment through the brain and agent and back again.
The framing invites a PyTorch analogy: DeepSoma positions itself as a common substrate for defining, training and running embodied intelligence, where the objects are physical environments and biological brains rather than tensors in a data center. The ambition is the interesting part, but the honest caveats matter just as much. DeepSoma is a young company, and the open questions are real — can the platform scale to larger whole-brain networks, how closely do its simulations track actual neural activity, how much of the brain-to-body mapping still leans on manual rules, and can cross-body transfer be reproduced against public benchmarks. This is a platform and a promise today, not a proven bio-equivalent brain.
vivo's BlueLM 3.5 Nano 3B topped the latest on-device model rankings, with a score closing in on mainstream cloud models. SuperCLUE's new OnDevice benchmark tests models that actually run locally inside a phone, using cloud models only as a reference line and excluding them from the final ranking, and vivo's 3-billion-parameter BlueLM took first overall with a composite score around 89.9. It's a small but meaningful data point in the race to move AI fully offline — where a model small enough to fit a handset beats bigger ones that can't, and where privacy, latency and running cost all improve at once.
What to watch: whether DeepSoma can reproduce its cross-body transfer against public benchmarks, and which phone makers rush to answer vivo's on-device lead.
Is whole-brain emulation a genuine research frontier, or an expensive demo that won't scale? Tell us in the comments.
