Acrab says a $1,000-ish mini PC can run frontier-scale AI locally. The public benchmark is on a smaller model.

A Singapore hardware startup called Acrab went public on July 23, 2026, with a launch that, if its claims hold, would redraw the floor for running frontier-scale AI on a desk. The company says its GΞLIX 1 system-on-chip, packaged in a small desktop called the Agent Box, can handle open-source models in the 100-billion-parameter class, the size that today usually lives behind a paid cloud API or a Nvidia DGX Spark workstation priced around $5,000. Acrab's pitch is that Agent Box will land at "roughly one-fifth" of that price and roughly half the power draw. Agent Box pricing has not been announced.

The claim rests on an architecture that borrows more from mobile silicon than from a data-center GPU. GΞLIX 1 is a system-on-chip, the kind of integrated processor that powers smartphones, and it pairs a 20-core Arm CPU with a multicore neural processing unit (NPU) tuned for transformer inference, plus a small 3 TFLOPS GPU and a memory subsystem capable of 273 GB/s of unified bandwidth, with an 8 MB L1 cache and a 768 GB/s L2 cache. The 5-nanometer die is rated at 700 TOPS, a trillion operations per second, a figure Acrab uses to position the part against Nvidia's entry-level workstation rather than its H100 or B200 accelerators. (Acrab press release via PR Newswire, July 23 2026)

The number Acrab is leading with is a prefill-rate benchmark, the speed at which the chip chews through the prompt before it starts generating tokens. Acrab says GΞLIX 1 hit 1,416.8 tokens per second on Gemma 26B A4B, a 26-billion-parameter mixture-of-experts model, with a 40K KV cache, a memory pool that stores the key-value pairs of the model's attention layers, and a 10,000-token input. On the same test, an Apple Mac Mini M4 Pro managed 188.9 tokens per second. Acrab frames that gap, about 7.5x, as evidence that Agent Box can stand in for Nvidia's box. (Wccftech coverage of the Acrab launch)

There is a gap between the test and the marketing. Acrab's launch targets 100-billion-parameter models, the size of Llama 3.1-405B-style frontier systems, but the only disclosed benchmark is on a 26B-class model, roughly a quarter of that scale. Decode ratio and memory pressure do not scale linearly: doubling model size usually quadruples memory traffic because the key-value cache grows with both depth and sequence length, so a 26B prefill result is suggestive rather than dispositive. A 100B-class run on the Agent Box, with the same KV-cache and input configuration, is the test that would either validate or break the claim. Acrab has not released one. (Unite.AI on the 100B-class target)

The backer question is the other live thread. Acrab's launch-video page carries a "Contemporary Amperex Technology" (CATL) copyright string in its footer, a trail first noticed by Japanese tech outlet PC Watch, which also noted that CATL is the world's largest maker of electric-vehicle batteries. Acrab has not named CATL as an investor, and no CATL press release, executive statement, or regulatory filing in the source set confirms a corporate stake. The English-language framing of Acrab as "backed by the world's largest battery maker" therefore rests on that single indirect signal. It is interesting context, not a confirmed relationship.

Acrab's financial base is better documented. The company emerged from stealth in June 2026 with more than US$350 million in cumulative financing; named backers include Vertex Ventures Southeast Asia & India, Vertex Growth, and K3, per DealStreetAsia. CEO Dr. Ken Phua is positioning the launch as a category move, from generative AI, where models answer questions, to "agentic" AI, where models drive actions: "Generative AI helped people find answers. Agentic AI will help them get things done." (Acrab press release)

The first independent review of Agent Box on a real 100B-class model, with disclosed hardware, KV-cache size, and prefill-versus-decode split, is the test that will tell readers whether the Acrab launch is the moment local AI crossed a price threshold, or a well-marketed product announcement leaning on a 26B benchmark to imply a 100B result.