#USA #Beijing #AI Research #Apex Intelligence #Apex-35B-V1.0

Apex Intelligence Launches Apex-35B-V1.0

Apex Intelligence has officially launched Apex-35B-V1.0, a cutting-edge general-purpose model developed through advanced automated AI research methodologies. Representing the forefront of AI technology, this model is not only built on the existing Qwen3.6-35B-A3B architecture but has also undergone rigorous mid-training and post-training phases to enhance its capabilities.

A New Era in AI Technology

Apex-35B-V1.0 stands out in its category, showcasing exceptional research and development (R&D) capabilities while maintaining a robust general-purpose application. According to internal evaluations conducted by Apex Intelligence, this model surpasses similar counterparts in key areas, demonstrating competitiveness against larger models on selected benchmarks.

In recent tests, Apex-35B-V1.0 exhibited outstanding performance across 21 varied benchmarks. Impressively, it outperformed the Qwen3.6-35B-A3B model in 20 tests and matched the performance of Recursive Flint (40B) on one benchmark while exceeding it on 18 others. These results underscore the model's versatility and enhanced capabilities in general-purpose tasks, specifically in coding, question answering, and executing complex workflows.

Automated AI Participation in Model Development

Unique to Apex-35B-V1.0 is the significant role played by AI itself throughout its development lifecycle. The system autonomously engaged in tasks ranging from data collection and experimental design to the execution of code and analysis of training results. This model is adept at executing end-to-end computational research workflows applicable in AI research, computational science, and robotic simulations.

In addition to its pioneering development process, Apex Intelligence has established Apex Research, an automated system aimed at enabling recursive self-improvement of AI models. This innovative approach involves identifying weaknesses, hypothesizing solutions, conducting experiments, and translating findings into measurable improvements. Noteworthy is how the system improved its performance on computational fluid dynamics (CFD) tasks by an impressive 32.1 points by crafting targeted training data aimed at rectifying previously identified failures.

Performance Optimization and Efficiency

Apex Research also plays a critical role in analyzing evaluation results to recognize patterns of failure and refining training methodologies accordingly. Specific adjustments have allowed for better handling of excessively long reasoning contexts and imbalanced data mixtures, leading to enhanced agent performance and training efficiency.

This feedback loop effectively creates an ongoing cycle of improvements that not only enhances the model's precision but also optimizes the infrastructure associated with its training and inference operations. As a result, Apex Intelligence reported 45-60% higher training throughput through various innovations in GPU load management and communication, while inference optimizations facilitate faster decoding times and lower token generation latency.

Availability and Collaboration

For those eager to experiment with this state-of-the-art model, the weights for Apex-35B-V1.0 are currently accessible on Hugging Face. Furthermore, accompanying code and deployment guidelines are provided on GitHub, allowing developers to implement this model in AI tasks, scientific computing, and other research initiatives. Detailed information regarding model versions, hardware prerequisites, and licensing can also be found on the respective project pages.

As Apex Intelligence continues its mission to evolve AI into a self-improving research collaborator, plans to develop a larger self-evolving foundation model are already in the pipeline. This forthcoming model will aim to enhance AI research capabilities and broaden applications within a variety of fields, including advanced chip design and quantum computing.

In summary, Apex-35B-V1.0 signifies a notable step towards creating AI that not only assists but actively contributes to scientific discovery and innovation. Given the rapid advancements and shifts in AI technology, the future holds exciting prospects for Apex Intelligence and its upcoming innovations.