Geely unveils Eva Cab purpose-built robotaxi prototype

Geely showcased the Eva Cab robotaxi prototype at the Beijing Auto Show on April 24, 2026, Reuters and multiple trade outlets reported. The vehicle is described in press coverage as a purpose-built, no-steering-wheel robotaxi developed with Geely's ride-hailing arm, CaoCao Mobility, and slated for production starting in 2027, with expanded delivery and fleet deployment reported for 2028 and a reported fleet target of 100,000 by 2030 (Reuters).

Press reports attribute an onboard AI stack using a 196-billion-parameter model and a compute subsystem delivering up to 1,400 TOPS of performance (Interesting Engineering; VNExpress). Coverage describes a hardware platform said to include two Nvidia Thor chips, an array of LiDARs and cameras totaling roughly 43 perception components in some accounts, and claims around inference throughput such as 350 TPS and faster-than-human reaction metrics in vendor reporting (Interesting Engineering; VNExpress; CarNewsChina).

Companies building purpose-designed robotaxi hardware typically integrate higher compute density and a larger, more diverse sensor suite than retrofit approaches, trading component cost for tighter software-hardware co-design. Industry observers note that onboard LLM-scale models in mobility stacks shift compute and thermal design constraints onto the vehicle, increasing emphasis on efficient accelerators and power management rather than purely on perception stacks.

Industry context: The Eva Cab coverage frames Geely's effort as part of a wider wave of OEMs and mobility platforms moving from retrofit testbeds to dedicated robotaxi platforms. For practitioners, that pattern changes operational trade-offs: fleet operators and OEMs weigh higher per-vehicle hardware cost against potential reductions in software complexity, interior optimisation for shared rides, and lifecycle maintenance. Integrating a large language-model-sized component into driving stacks also raises the bar on validation, simulation fidelity, and interpretability requirements for production safety cases.

Editorial analysis: Observers should monitor independent verification of the performance claims reported in vendor and trade coverage, uptake and pilot results from the named initial markets (Reuters listed Abu Dhabi, Hong Kong and several Chinese cities), and how Geely/CaoCao publish technical validation data (for example disengagement rates, scenario coverage, and third-party audits). Also watch component supply lines: high onboard compute claims place sustained demand on automotive-grade accelerators such as Nvidia Thor and on high-resolution LiDAR suppliers.

For ML engineers and AV systems teams, purpose-built vehicles like the Eva Cab underline the need to design across domains: model architecture, real-time inference stacks, sensor fusion latency budgets, thermal and power constraints, and fleet telematics. Industry readers should treat vendor performance claims as starting points for technical evaluation rather than conclusive proof; independent testing and published metrics remain essential.

The preceding reported facts are drawn from Reuters, Interesting Engineering, VNExpress, Quattroruote and other trade reporting compiled at the Beijing Auto Show. Editorial paragraphs are LDS analysis and framed as industry-level observations, not as statements about Geely's internal intentions or future actions.