As AI moves beyond the data centre and into wearables, vehicles, industrial equipment and other edge environments, a fundamental question emerges: does every AI system need a connection to the cloud?

SLM360X, developed by 360 Labs' CTO Saurabh Kumar, explores a different approach: a compact, multimodal foundation model designed to operate entirely on-device, including in denied, degraded, intermittent or limited-bandwidth environments.

Designed for the edge

SLM360X combines several specialised components within a single architecture. These include a routed Mixture of Tiny Experts (MoTE), on-device retrieval augmentation, a generative decoder with an inline tool runner, and multimodal inputs covering speech, vision and acoustic classification.

The system is designed around a simple principle: where connectivity, latency, privacy or operational resilience are critical, AI should not necessarily depend on a remote inference service.

The model and its supporting components can instead be deployed directly onto the hardware an operator already carries, from smart glasses and body cameras to rugged tablets, industrial helmets and embedded systems.

The MATSYA deployment

The architecture was developed and evaluated through MATSYA, a naval deployment designed around real-world operational requirements.

The resulting system comprises a 429 MB model bundle, with the complete runtime remaining below approximately 650 MB when supporting components such as speech recognition and text-to-speech are included.

The evaluation reported:

98.5% router accuracy across seven specialist domains;
approximately 3 seconds end-to-end voice-in/voice-out on a general-purpose CPU;
deterministic inference, with repeated inputs producing byte-identical outputs under the defined inference contract; and
operation without a network connection.

The system also incorporates an on-device corpus containing operational and doctrinal material, allowing retrieval to take place locally rather than requiring information to be sent to an external model provider.

Why determinism matters

For many consumer AI applications, variability in model outputs is acceptable. In operational environments, it can be a materially different proposition.

SLM360X therefore treats determinism as an architectural requirement rather than simply a desirable characteristic. The system uses fixed inference conditions and deterministic tooling so that identical inputs can produce reproducible outputs.

This also creates the basis for more auditable AI systems, where the model, underlying corpus and tool execution can be versioned and retained locally.

Beyond defence

Although the initial architecture was shaped by requirements of a specific deployment in a defence sector use case, the underlying approach is not limited to defence.

The paper identifies potential applications including:

smart glasses for field medics;
first-responder body cameras;
industrial safety helmets;
maritime bridge systems;
UAV operator devices;
warehouse and industrial AR systems;
embedded drone systems; and
consumer wearables operating on sensitive personal data.

Across these environments, the common requirement is not necessarily maximum model capability. It is reliable intelligence within a constrained operational environment.

Edge and cloud as complementary architectures

SLM360X is not positioned as a replacement for frontier cloud models.

Rather, it demonstrates a different point in the AI architecture: a smaller, specialised model can provide a dependable baseline of capability when network access is unavailable or undesirable, while a cloud-based model can provide greater open-ended capability when connectivity permits.

In this sense, the edge can provide the floor; the cloud can provide the ceiling.

For organisations deploying AI into increasingly physical and operational environments, the question may therefore be less about choosing between edge and cloud, and more about determining which parts of the AI workload should be capable of operating independently of the cloud.