On July 29, Singapore-based Acrab – founded by former Arm China co-CEO Pan Zhenyuan – launched its 5nm edge AI chip Gelix 1, targeting local LLM inference on personal devices with a peak 700 TOPS of AI compute.

The chip features a heterogeneous architecture integrating 20-core Arm CPU, 3 TFLOPS GPU, dedicated NPU, and AV processing engine. It delivers 8MB L1 cache, 768GB/s shared L2 bandwidth, and 273GB/s unified memory bandwidth – designed to handle 100-billion-parameter models entirely on-device, without cloud dependency.

In internal testing, running a Gemma 26B model with a 40K token context, Gelix 1 achieved 1,416 token/s pre-fill rate – 7.5x faster than Apple's M4 Pro in the Mac Mini under the same conditions.

Acrab has also built a full software stack, including agent reference designs and the Agent Box personal edge AI system. The company, founded in 2024, has raised over $350 million from investors including Huangpu River Capital and Temasek's Vertex, according to public sources.

The chip's pricing and launch timeline have not been disclosed, and all performance data is internal – yet to be independently verified.

From ICgoodFind: 700 TOPS in a Mac Mini-sized box – that's a new category. Gelix 1 doesn't just compete with cloud GPUs; it makes local 100B-parameter AI a desktop reality for the first time.