Nine Zenith · Zenith-Act industry large model completes a new round of upgrade

In June 2026, the Zenith-Act industry large model of NineZenith completed a major upgrade. This upgrade has achieved significant breakthroughs in three dimensions: inference efficiency, industry knowledge injection, and multi-modal understanding, further consolidating the competitiveness of NineZenith's technology in the field of industry large models.

Core Upgrade

MoE architecture fully upgraded

The new version of Zenith-Act large model has switched from a dense architecture to the MoE (Mixed Expert) architecture, keeping the total parameter count within a reasonable range, but reducing the active parameter count per inference to 37B. This change has brought significant performance improvements:

Deep injection of industry knowledge

This upgrade adds over 8 million pieces of industry professional knowledge data, covering four vertical fields of government affairs, energy, manufacturing, and people's livelihood. Through three-stage training of continuous pre-training (CPT) + instruction fine-tuning (SFT) + reinforcement learning (RLHF), the model has achieved a significant improvement in the accuracy of professional question answering:

• Government Q&A: Accuracy improved from 85.3% to 93.1% (based on an internal evaluation set of 5000 questions)
• Energy equipment fault diagnosis: Accuracy improved from 78.6% to 92.4% (based on actual data from a certain energy group)
• Manufacturing process optimization suggestions: Expert adoption rate increased from 67% to 89%
• Interpretation of People's Livelihood Policies: Semantic understanding accuracy improved from 88.1% to 94.7%

Multimodal understanding enhanced

The new version of Zenith-Act has enhanced its visual understanding capabilities, supporting automatic recognition and analysis of industrial site photos, equipment monitoring images, and urban governance scene images. For example, in the energy inspection scenario, the model can automatically identify defects such as hot spots and hidden cracks by analyzing the photos of photovoltaic panels taken by drones.

Technical Details

Training Data Governance

The key to industry knowledge injection is not the amount of data, but the quality of the data. NineZenith technology has established a complete industry data governance pipeline:

1. Data Collection: From desensitized acquisition from cooperative units, public policy documents, industry technical standards, professional textbooks, etc.
2. Data Cleaning: Duplicate removal, noise reduction, format standardization, and automatic filtering of low-quality content
3. Expert Annotation: Industry experts participate in the manual verification and labeling of key samples
4. Continuous Iteration: Continuously supplement high-quality data based on actual usage feedback

Inference Optimization

In addition to the efficiency improvement brought by the MoE architecture itself, Zenith-Act also adopts the PagedAttention technology of the vLLM inference engine, which greatly reduces the GPU memory fragmentation of KV caching. Combined with quantized inference (INT8), the number of concurrent requests supported by a single card can be increased by 4 times.

Future Plans

In the next phase, the Zenith-Act large model will focus on continuously optimizing in the following directions:

• Extended context support (target 128K→256K), compatible with complex policy and regulation analysis scenarios
• Edge-end lightweight version, meeting the offline inference needs of data not leaving the domain
• Enhanced tool calling capabilities, more accurately connecting to existing business systems of governments and enterprises

NineZenith technology has always adhered to the orientation of industry needs, continuously refining the professional capabilities of the Zenith-Act large model, making AI truly a good helper for government and enterprise customers.