The integration lets Virtuals split a 744 billion parameter model across multiple GPUs, a key step toward running frontier-scale AI in decentralized environments
Running a model with roughly 744 billion parameters is not something you do on a single graphics card. Virtuals Protocol just partnered with Leyten to make sure it doesn’t have to.
The AI agent platform has integrated Leyten’s shard engine, a system designed to distribute large-model inference across multiple GPUs over a network. The immediate target: running GLM-5.2, the open-weight model from Z.ai that dropped publicly under an MIT license on June 16, 2026. The combination gives Virtuals a path to frontier-scale AI inference without relying on centralized cloud providers or single massive GPU clusters.
What GLM-5.2 actually is, and why it matters here #
GLM-5.2 is a big model. We’re talking approximately 744 billion total parameters, though only around 39 to 40 billion are active per token. In English: the model uses a mixture-of-experts architecture that keeps most of its knowledge stored but only fires up a fraction of it for any given task, keeping compute costs manageable despite the enormous overall size.
The model also ships with a context window of 1 million tokens. That’s five times larger than its predecessor, GLM-5.1.
Z.ai released GLM-5.2 to subscribers on June 13, 2026, before making it publicly available three days later. The MIT license means anyone can use, modify, and deploy it commercially.
How Leyten’s shard engine solves the hardware problem #
Leyten built a different approach. Its shard engine uses pipeline-parallel inference, which essentially slices a large model into pieces and distributes those pieces across separate GPUs connected over a network. No single node needs to hold the entire model in memory.
Where Virtuals Protocol fits in the AI agent landscape #
Virtuals Protocol operates in the AI agent vertical of crypto, specifically focused on the creation and monetization of onchain AI agents. These are autonomous digital entities that can transact, execute tasks, and interact with blockchain protocols. The ecosystem runs on its native token, VIRTUAL.
GLM-5.2 has been positioned as competitive with proprietary frontier models but at significantly lower operational costs. A model with 1 million tokens of context and strong agentic coding capabilities is precisely the kind of foundation that makes autonomous agent behavior plausible rather than aspirational.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our