Apex Intelligence releases Apex-35B-V1.0, cites 32.1-point CFD-Bench gain

Apex Intelligence, a Chinese AI startup founded in June 2026, released its first model, Apex-35B-V1.0, on October 28 and made the weights available on Hugging Face. The 35-billion-parameter model is built on Alibaba’s Qwen3.6-35B-A3B and was refined through mid-training and post-training for coding, agent workflows, general tasks, and research in AI, computational science and robotics simulation.

Apex Intelligence highlights its automated research system, Apex Research, which generated training data, ran experiments, designed the training plan and optimized GPU kernels. The company says the system identified weaknesses on computational fluid dynamics tasks, produced new data, and lifted the model’s CFD-Bench score by 32.1 points. Reported AI-driven optimizations increased training throughput by 45-60 % and made decoding 1.6-1.8 times faster.

Internal benchmarks claim Apex-35B-V1.0 beats Qwen3.6-35B-A3B and Recursive SuperIntelligence’s Flint on six research tests, exceeds GLM-5.3 on FrontierScience and PaperBench, and surpasses DeepSeek-V4.1-Flash-1017 on the human-difficult split of BioMysteryBench. On robotics tests RoboRSI and Real2Sim it scores 44.5 and 21.55, respectively (all benchmark results are from the company’s own evaluation), close to GLM-5.3-Flash’s 52.4 and 21.4 (company’s own evaluation).

Real2Sim asks a model to rebuild a working physics simulation from a single photograph: identify the objects, choose simulation assets with similar visual and physical properties, and recreate their layout so robot actions can be tested. RoboRSI measures how well a model controls simulated robots by writing code and revising it based on visual feedback.

The company was founded by Chen Yongchao, an assistant professor at Tsinghua University’s College of AI. He frames the work as recursive self-improvement, the idea of AI systems helping to build their successors. He told Tencent Technology that AI has made clear progress in carrying out research but still lacks research taste, the judgment of which problems are worth pursuing.

Our read: The 32.1-point CFD-Bench gain stems from an automated loop that creates targeted training data and tunes hardware, a workflow that could shorten model-development cycles. Reported throughput and decoding speed improvements suggest lower compute costs, but all results are internal; external validation will be required to gauge true performance relative to peers.