While the AI industry remains obsessed with the "parameter race," Weibo AI has delivered an unexpected breakthrough, charting a new path in the large model arena.

Recently, Weibo officially launched its first self-developed open-source large model, VibeThinker. This lightweight contender, with only 1.5 billion parameters, outperformed DeepSeek R1—a model with 671 billion parameters, hundreds of times larger—in top-tier international mathematical benchmark tests.

Even more striking is its post-training cost of just $7,800, slashing expenses by dozens of times compared to models like DeepSeek-R1 and MiniMax-M1. This breakthrough not only redefines technical evaluation standards for large models but could also shift the AI industry from a "scale competition" to an "efficiency revolution."

**Industry Disruptor: Small Model Challenges Parameter Worship** Historically, parameter count has been viewed as the core metric for AI model capability. The consensus was that complex reasoning required models with over 100 billion parameters, while smaller models were dismissed as inherently limited.

But can a compact model unlock hidden reasoning power through innovative training strategies? Weibo’s VibeThinker answers with a resounding "yes."

While most AI firms adhere to the "bigger is smarter" Scaling Law, Weibo’s researchers optimized model architecture and training paradigms, introducing the "Spectrum-to-Signal Principle" (SSP) method. The result? A 1.5B-parameter lightweight model that defeats giants hundreds of times its size.

VibeThinker’s release sparked global attention after acing rigorous benchmarks in mathematics and coding: - HuggingFace featured its research paper. - On AIME24, AIME25, and HMMT25 math test sets, it surpassed DeepSeek-R1-0120 (671B parameters) and rivaled MiniMax-M1 (456B), even matching Gemini 2.5 Flash and Claude Opus 4. - In LiveCodeBench v6 (programming algorithms), it equaled models like Minstral.AI’s Magistral-Medium-2506, despite their far larger scale.

VibeThinker proves that sophisticated algorithm design can enable small models to match—or surpass—behemoths in complex reasoning tasks, reshaping cost structures and technical roadmaps for the AI industry.

*Note: The current experimental version focuses on advanced math and competitive programming, not general chat applications.*

**Cost Revolution: $7,800 Training Reshapes Industry** Training costs have long hindered AI adoption, but VibeThinker’s cost-performance ratio is transformative.

Industry benchmarks show post-training costs often exceed hundreds of thousands of dollars: - MiniMax’s M1 (June 2025) required 512 H800 GPUs for three weeks at ~$535,000. - DeepSeek’s R1 (September 2025) cost $294,000 (excluding $6M foundational LLM development).

In contrast, VibeThinker’s entire post-training (SFT + RL phases) consumed just 3,900 GPU hours—totaling $7,800 at market rates. This 30–60x efficiency leap democratizes high-performance AI, enabling SMEs, labs, and universities to innovate.

**Deployment: Weibo’s AI Ecosystem Expands** Weibo is aggressively integrating AI across its platform: - Its proprietary "Zhiwei" LLM (launched 2024) powers features like *Weibo Smart Search* (50M MAU) and *Comment Robert*, an AI interactive account with 2M fans.

With VibeThinker, Weibo plans to leverage its unique vertical data (e.g., psychology) to build models that better understand public sentiment and social needs. The breakthrough could also slash operational costs for real-time AI services, fueling a smarter, more dynamic social ecosystem.

VibeThinker’s efficiency may soon enhance core products like Smart Search, potentially birthing a new "social super-ecosystem" blending connectivity and AI-powered intelligence.