Skywork AI

Skywork AI sells large-scale foundation-model software and API access to its “Skywork” family of bilingual (Chinese-English) language and multimodal models. Pricing follows a pay-as-you-go token model for developers and negotiated enterprise licenses for on-premise or private-cloud deployments, placing the offer in the mid-range to premium tier compared with open-source weights. All sales and technical onboarding are handled online through the company’s developer portal and direct sales team; no retail channel exists. The brand’s core edge is a training corpus of 1.2 trillion high-quality Chinese tokens plus 300 billion English tokens, yielding state-of-the-art scores on C-Eval, CMMLU and MMLU benchmarks without U.S. export-controlled hardware. Skywork open-weights releases (7B, 13B, 40B) under Apache 2.0 have been downloaded more than 1.3 million times in six months, creating viral awareness among AI researchers. A forthcoming 180B sparse mixture-of-experts model is already in closed beta with major fintech and telecom clients. Primary buyers are Chinese tech companies, SaaS vendors, and government institutions that need domestically hosted, compliant generative-AI capability for customer service, content creation, and code generation. Secondary customers include global AI researchers who want bilingual benchmarks and reproducible open models. Users value data sovereignty, Mandarin linguistic fidelity, and transparent model cards that meet China’s pending algorithmic filing rules. Skywork AI competes with both U.S.-based closed APIs and domestic Chinese lab models. It differentiates by releasing fully open weights at competitive sizes, publishing complete training logs, and offering on-premise deployment packages that keep data inside mainland infrastructure, satisfying regulatory requirements foreign providers cannot match.