Acrab Unveils GELIX 1 Edge AI SoC

Acrab's published benchmark used Google's Gemma 26B A4B model at a 40,000-token context length. The company reported a prefill rate of 1,416.8 tokens per second for GELIX 1, compared with 188.9 tokens per second for an M4 Pro Mac mini in the same test, a roughly 7.5x difference. Wccftech reported those figures and noted that independent verification had not yet occurred.

The comparison therefore should not be treated as a settled performance ranking. Prefill throughput measures the processing phase that ingests a prompt, and it is distinct from decode throughput, which governs token-by-token generation speed. Hardware comparisons can also vary substantially with model quantization, batch size, prompt length, runtime implementation, thermal limits, and memory capacity.

Acrab has not disclosed GELIX 1's memory capacity in the materials reviewed. That omission matters for the company's 100-billion-parameter model claim: whether such a model fits and performs acceptably locally depends on its precision or quantization, the memory allocated to runtime and key-value cache, and the target context window. SSPAI similarly reported the 100-billion-parameter support claim while noting that the specification available did not provide a memory-capacity figure.

In a direct statement released through PRNewswire, Acrab CEO Dr. Ken Phua said, "Generative AI helped people find answers. Agentic AI will help them get things done." He added that running 100-billion-parameter-class models in a desk-sized system presents a significant computing challenge.

According to the same release, Acrab frames local execution as a way to reduce response latency, retain sensitive data under user control, and keep core functions available when cloud connectivity is limited. Hardware-News.de also describes the platform as combining local LLM inference with multimodal models, tool and device integration, and persistent-memory functions.

For ML practitioners, the announcement underscores a broader edge-compute pattern: model size alone is an incomplete measure of deployability. Viable local agent systems require enough memory for weights and context, sustained bandwidth, an optimized inference stack, and reproducible performance data across both prefill and decode workloads. Acrab has not published pricing, availability, or independently verified performance results in the cited coverage, leaving those details as important open questions for teams evaluating the platform.