Singapore startup Acrab announced its GELIX 1 edge AI system-on-chip and Agent Box reference system on July 23, targeting local inference for models in the 100-billion-parameter class. Acrab says the 5 nm chip combines a 20-core Arm CPU and multicore NPU with 273 GB/s unified-memory bandwidth; independent reports describe its 7.5x M4 Pro result as an unverified company benchmark.
Singapore startup Acrab announced its first edge AI system-on-chip, GELIX 1, alongside a compact reference system called Agent Box, on July 23. Digitimes reports that the chip was unveiled on July 22; Acrab's official release is dated July 23. The company is targeting local execution of large language and multimodal models, including models in the 100-billion-parameter class, rather than continuous reliance on cloud infrastructure.
According to the company-issued PR Newswire release, GELIX 1 is fabricated on a 5 nm process and combines CPU, GPU and neural-processing resources in a unified-memory architecture. Acrab lists a 20-core Arm CPU, a multicore NPU and 273 GB/s of unified-memory bandwidth. It describes Agent Box as a personal edge AI system for local large-model inference, persistent memory, multimodal interactions and agent orchestration.
Performance claims need independent testing
Acrab's benchmark used Google's Gemma 26B A4B model with a 40K KV cache and a 10,000-token input. 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.
Digitimes describes the comparison as Acrab's claim. Wccftech reported the same figures and, citing Digitimes, said independent verification had not yet occurred. The result therefore should not be treated as a settled performance ranking. Prefill throughput measures the prompt-ingestion phase and is distinct from decode throughput, which governs token-by-token generation. Comparisons can also shift with quantization, batch size, prompt length, runtime, thermals and memory capacity.
Neither Acrab's release nor the retrieved independent reports state GELIX 1's memory capacity. That omission matters for the 100-billion-parameter claim: whether a model fits and performs acceptably depends on weight precision, runtime overhead, KV-cache allocation and the target context window.
Local inference and Agent Box
Acrab CEO Ken Phua said running 100-billion-parameter-class models in a desk-sized system is a significant computing challenge. The company frames local execution as a way to reduce latency, retain sensitive data under user control and keep core functions available when connectivity is limited.
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 results across prefill and decode workloads. Acrab has not published pricing, availability, memory capacity or independently verified performance in the retrieved materials. Those details remain the key questions for teams evaluating the platform.
Key Points #
- 1Acrab announced GELIX 1 and Agent Box for local large-model inference, but commercial specifications remain incomplete.
- 2The reported 7.5x M4 Pro prefill result is company benchmark data without retrieved independent verification.
- 3Memory capacity, quantization, context length and decode speed will determine whether 100-billion-parameter models are practically deployable.
Scoring Rationale #
GELIX 1 is a notable edge-inference hardware announcement because it targets local execution of very large models and agent workloads. Its practical relevance remains constrained by undisclosed memory capacity, pricing, availability and the absence of independently verified performance results.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.