Researchers at Texas A&M University have introduced EMMI, a framework designed to optimize Multimodal Large Language Model (MLLM) inference at the edge. By performing cross-modal fusion and learned compression locally, the system reduces communication payloads by 32x and improves end-to-end latency by up to 3.4x in bandwidth-constrained environments.
The EMMI framework addresses the primary barrier to deploying high-capacity multimodal models on resource-constrained hardware by shifting the communication boundary. By fusing vision and text data into a compact latent representation before transmission, the system bypasses the need to send raw sensor streams or heavy intermediate activations. For the streaming and IoT ecosystem, this enables sophisticated AI reasoning on low-bandwidth links where cloud-only processing was previously too slow. The decoupling of compression from the server-side LLaVA backbone allows for flexible updates to edge hardware without retraining central models. Watch for future evaluations on specific mobile NPUs to determine real-world power efficiency gains.
The push to run multimodal AI models closer to the network edge is accelerating across the telecom and streaming sectors, where latency and bandwidth constraints make cloud-only inference impractical. In June 2026, Ericsson launched its AI in RAN commercial software subscription, claiming up to 20% higher downlink throughput and up to 10% better spectral efficiency across more than 15 live deployments, demonstrating that edge-side AI processing can deliver measurable performance gains at scale. This commercial momentum mirrors the academic work behind EMMI, which targets the same fundamental problem of reducing data movement between edge devices and central compute resources.
Nokia has taken a complementary approach by building agentic AI directly into its mobile core network products, positioning edge inference as a commercial differentiator for operators. Nokia is deploying generative AI and agentic technologies for root cause analysis and autonomous decision-making at the network edge, reducing call setup times from roughly 10 seconds to one or two seconds in some use cases. The company has also introduced a "Mobile Core Early Access" program that lets operators trial AI-based features before full deployment, a go-to-market model that could inform how academic frameworks like EMMI eventually reach production environments. Meanwhile, Nokia announced partnerships with AWS and Databricks to build a unified data and cloud control layer for autonomous networks, claiming operators are already achieving automation rates above 90% and service delivery times under four hours.
The competitive landscape for edge AI inference is further shaped by diverging architectural strategies among major vendors. Ericsson and Nokia are taking fundamentally different paths on AI-RAN, with Nokia building its entire Layer 1 RAN strategy around Nvidia's CUDA-accelerated compute platform following a $1 billion investment from the chipmaker. This hardware-level divergence has direct implications for frameworks like EMMI, which must eventually target specific edge silicon to demonstrate real-world power and throughput gains. Nokia is using AI-RAN to improve spectral efficiency while creating an evolutionary path from 5G-Advanced toward AI-native 6G, with T-Mobile US testing both vendors' approaches. For streaming applications, the ability to run lightweight multimodal models on edge hardware without saturating backhaul links represents a practical path toward real-time content understanding, quality monitoring, and adaptive encoding at the point of capture. Recent GPU-accelerated edge AI video analytics further highlight how these platforms are evolving to handle complex visual data locally.
Texas A&M University researchers developed the EMMI framework to optimize edge MLLM inference. By performing cross-modal fusion and compression locally, it reduces communication payloads by 32x and improves end-to-end latency by 3.4x. This innovation enables sophisticated AI reasoning on bandwidth-constrained devices, bypassing the need for heavy raw data transmission.
EMMI is a framework developed by Texas A&M University researchers designed to improve edge MLLM inference by performing cross-modal fusion and learned compression locally on edge devices.
The EMMI framework reduces communication payloads from 8,192 bytes to 256 bytes, representing a 32x reduction.
In 0.1 Mbps IoT environments, EMMI improved end-to-end latency by 3.4x, reducing processing time from 903ms to 269ms.
The TaskAwareAE compression strategy used in EMMI maintains accuracy within 0.16 percentage points of uncompressed baselines.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source