How the 32B ‘Zhiyu’ Model is Revolutionizing Intelligent Operations

Amid the wave of digital transformation, AI is reshaping industries, and intelligent operations (SRE) are becoming increasingly vital. To address challenges such as insufficient answer accuracy, weak executability, and poor scenario adaptation, the Stability Lab introduced the groundbreaking SRE‑domain large model Zhiyu , injecting new vitality into intelligent operations.

1. Model Overview: A 32B Knowledge Engine

Zhiyu is a 32‑billion‑parameter SRE‑domain LLM built on Qwen3‑32B, refined through incremental pre‑training and post‑training (including fine‑tuning and reinforcement learning). It inherits the base model’s capabilities while delivering superior domain expertise and practicality.

2. Core Highlights: Triple Enhancement of Knowledge, Skills, and Determinism

Domain Knowledge Learning The model absorbs extensive SRE public knowledge, including books, papers, industry reports, and whitepapers such as the “SRE Practice Whitepaper,” establishing a solid knowledge base.

Professional Skill Enhancement By learning from human‑operated experience, component documentation, real fault cases, and platform‑derived data, Zhiyu markedly improves SRE workflow, tool usage, and specialized skills like SQL/PromQL generation.

RAG Determinism Boost Advanced data construction and governance mechanisms ensure accurate, reliable model outputs, strengthening operational decision‑making.

3. Data Construction: Diverse, High‑Quality Training Sets

Zhiyu’s training data spans public domain knowledge, human operational experience, generic component documentation, real fault cases, platform‑extracted data, and reinforcement‑learning‑generated samples. A full‑pipeline toolchain (production/collection, governance, filtering/enhancement) guarantees timeliness and quality.

Key sources include:

Public SRE books, papers, reports, and whitepapers.

High‑quality operational experience from SRE experts.

Documentation of generic components, especially domestic ones.

Comprehensive fault case data covering summary, background, trajectory, root cause, actions, and optimization.

Real operational data extracted from enterprise platforms.

Reinforcement‑learning‑generated data from simulated SRE tasks.

4. Model Training: Three‑Stage Progressive Advancement

The training process consists of incremental pre‑training, fine‑tuning, and reinforcement‑learning post‑training. Incremental pre‑training and supervised fine‑tuning embed deep domain knowledge; subsequent fine‑tuning and reinforcement learning enhance professional skills; finally, DPO‑based reinforcement learning achieves a qualitative leap in general capabilities.

5. Evaluation & Deployment: Outstanding Performance, Flexible Integration

Using an independently built test set covering SRE knowledge, human operational experience, component usage, and professional skills (30,161 items), Zhiyu outperforms the base model across all dimensions. It supports NVIDIA VLLM and Ascend VLLM deployment options, with detailed hardware specs and launch commands for rapid, efficient deployment.

6. Usage & Open Source: Easy Access, Collaborative Future

Zhiyu can be accessed via OpenAI‑compatible and ModelScope APIs, with comprehensive code examples provided. Released under the Apache 2.0 license, it encourages enterprises and developers to contribute to the SRE‑domain training dataset, fostering collective advancement of intelligent operations.