90B-Param AI Model Runs Offline on Smartphone with New Efficiency Tech

The demonstration of a 90B-parameter model running locally on a smartphone suggests a future where edge AI doesn't require constant cloud offload—slashing latency and privacy risks. For AI developers, the HXS layer could unlock on-device inference for the largest open models, a leap from today's 7B-13B limits.

PALO ALTO, Calif. — Trust Carbon Infrastructure, operating under Zenith Flow Innovations LLC, claims a breakthrough that could reshape the economics of AI serving. The company announced signed measurements for its patent-pending HXS compute-efficiency layer, showing that it reduces energy consumption on existing NVIDIA H100 GPUs by up to 50.5% while simultaneously boosting throughput by 167%, all without altering model outputs. The findings, released via GlobeNewswire, come as data center operators grapple with skyrocketing energy demands driven by the proliferation of large language models and AI vision systems.

“

Under a standard configuration across five widely used open-source models — Alibaba's Qwen2.5-72B-Instruct, Meta's Llama 3.3 70B, DeepSeek's R1-Distill 70B, Google's Gemma 3 27B, and Microsoft's Phi-4 — energy use fell between 17.2% and 27.4%.

The testing was conducted on a single NVIDIA H100 80GB graphics card using the vLLM 0.26.0 serving stack. Under a standard configuration across five widely used open-source models — Alibaba's Qwen2.5-72B-Instruct, Meta's Llama 3.3 70B, DeepSeek's R1-Distill 70B, Google's Gemma 3 27B, and Microsoft's Phi-4 — energy use fell between 17.2% and 27.4%. Peak GPU operating temperatures dropped by up to 15.0 degrees Celsius, easing cooling burdens. More striking was a deeper optimization profile applied to Microsoft's Phi-4: power draw plummeted from 306 watts to 152 watts, a 50.5% reduction, while throughput soared from 109.8 to 293.7 requests per second — a 167% increase — at the same latency target, with no degradation in output quality.

The efficiency gain, if representative at scale, carries profound implications. AI inference already accounts for a significant and rapidly growing share of cloud electricity consumption. A 50.5% energy reduction per card translates directly into lower operational costs and a smaller carbon footprint for hyperscale data centers. The temperature drop further promises reduced cooling requirements, which often represent a third or more of total facility power. For an industry under pressure to meet sustainability targets without throttling innovation, HXS offers a rare no-compromise proposition: more performance from existing hardware while cutting power and heat.

Beyond the server room, Trust Carbon showcased the layer's edge capabilities. At an AI glasses event in New York on July 29, 2026, company representative Brayon Michael Pieske demonstrated a 90-billion-parameter AI vision model running entirely offline on a smartphone — a feat that normally requires server-grade GPU infrastructure. This suggests HXS can shrink the compute footprint of the largest models enough to enable on-device inference, with implications for privacy, latency, and accessibility of frontier AI.