Infinigence AI Leads NVIDIA GPU Benchmark as Chinese Firms Push Proactive AI

NVIDIA Research launched SOL-ExecBench on March 19, 2026, according to the report. SOL stands for Speed of Light and refers to theoretical performance boundaries derived from hardware capability. The benchmark extracts 235 operator optimization tasks from 124 real and frontier AI models covering language, vision, speech, video and multimodal domains, and divides them into four tracks: L1 basic operators, L2 fused operators, Quant quantization operators and FlashInfer-Bench inference operators. Unlike benchmarks that compare speedup against a software baseline, SOL-ExecBench uses a SOL Score tied to hardware upper limits and runs tests on high-performance compute cards.

As of Oct. 26, Infinigence AI ranked first in total score in L1 and Quant and second in L2 and FlashInfer-Bench, QbitAI reported. In L2, it took 39 of 82 first places. The competition includes Databricks, Y Combinator-backed SF Tensor, Cursor, Recursive, doubleAI and Geometric, according to the report. SF Tensor said on Sept. 25 that 583 teams had submitted more than 31,000 evaluations, averaging one public submission about every eight minutes.

Infinigence AI said it built an operator-optimization agent around SOL-ExecBench that links code generation, execution verification, performance analysis and experience accumulation. The system uses container isolation to separate the agent workspace, evaluation area and debugging area; skills and shared memory provide GPU programming knowledge and reuse of validated strategies; and optimization policies differ for compute-bound, memory-bound, fused-operator and small-shape tasks. The company says the work tests whether token consumption by agents can be converted into verified computing-performance gains and explores a path toward recursive self-improvement in AI infrastructure.

In a separate interview with Leiphone, Liu Rui, founding director of Huawei's Xiaoyi large-model application lab, said AI assistants need to move beyond the chat box from passive response to proactive service. Liu graduated from CUHK in 2022 and joined SenseTime before joining Huawei in 2024. At Huawei, he led conversational photo editing that was introduced at the 2025 Mate 80 launch, and the lab was established at the end of 2025, covering large language models, multimodal intelligence, agents and embodied intelligence. The lab released the Aura streaming video-understanding model in April, which supports real-time video interaction, proactive service and one-question-multiple-answer, according to the interview.

Also on Oct. 28, ifanr published a test of iFlytek Input 16.0.1, saying the keyboard's third-generation voice input "intent drafting" and personalized AI keyboard use task-level information cards and long-term memory to generate drafts from short spoken instructions across chat apps. The report said users can correct addresses and names by voice during input, with erroneous phrases removed after the key is released. iFlytek says the Spark-ASR-2.0 speech-recognition model supports 202 dialects and accents, raises dialect accuracy to 91% from 85%, high-noise accuracy to 84% from 66%, and English slot accuracy in mixed Chinese-English speech to 80% from 33%. The company also says more than 150,000 creators have built 300,000 expression skills, and over 60% of users have participated in shaping their own keyboards.