Breaking News: VibeThinker-1.5B Surpasses DeepSeek-R1 with $7,800 Post-Training Budget

In the latter part of 2025, a significant breakthrough emerged from a Chinese company in the field of open-source artificial intelligence.

Weibo’s AI division, a prominent Chinese social networking company, recently unveiled its latest creation, the VibeThinker-1.5B—a 1.5 billion parameter large language model (LLM) that is a refined version of Alibaba’s Qwen2.5-Math-1.5B, a competitor in the Chinese tech industry.

The VibeThinker-1.5B model is now accessible for free download and utilization by researchers and enterprise developers, including commercial applications, under the permissive MIT License on platforms like Hugging Face, GitHub, and ModelScope. A detailed technical report is also available on the open access science publishing site arxiv.org.

Despite its relatively compact size, VibeThinker-1.5B has achieved remarkable performance in reasoning tasks related to mathematics and coding, surpassing models many times larger than itself. It even outperformed the widely acclaimed R1 model from DeepSeek, a Chinese competitor with 671 billion parameters, on a formal reasoning benchmark.

Moreover, VibeThinker-1.5B has shown superior performance compared to other models such as Mistral AI’s Magistral Medium, Anthropic’s Claude Opus 4, and OpenAI’s gpt-oss-20B Medium, all while requiring significantly less infrastructure and investment.

What sets VibeThinker-1.5B apart is its cost-effectiveness in post-training, with a budget of only $7800 USD for compute resources, a fraction of what is typically required for similar-scale models.

It’s important to note that the development of large language models (LLMs) like VibeThinker-1.5B is carried out in stages, beginning with pre-training to establish language fluency and general knowledge before moving on to post-training, which involves teaching the model to reason through problems and align with human expectations using expert-written answers.

The Spectrum-to-Signal Training Approach

VibeThinker-1.5B’s exceptional performance can be attributed to its unique training framework, the Spectrum-to-Signal Principle (SSP).

Unlike traditional models that focus solely on single-answer correctness, SSP separates supervised fine-tuning (SFT) and reinforcement learning (RL) into distinct phases with different objectives:

• SFT (“Spectrum Phase”): Maximizing diversity in potential correct answers to enhance the model’s overall performance.
• RL (“Signal Phase”): Implementing a reinforcement learning system (MGPO) to identify and reinforce the most accurate solutions from the diverse pool of answers.

This approach allows smaller models to explore reasoning space more effectively and achieve signal amplification without the need for extensive parameter counts.

VibeThinker-1.5B challenges the conventional belief that larger models are the only path to improved reasoning performance. By prioritizing diversity in training, WeiboAI has demonstrated that smaller, more accessible models can compete with and even surpass billion-dollar systems in logic-intensive tasks.

Performance Comparison Across Domains

Despite its modest size, VibeThinker-1.5B excels in cross-domain reasoning, outperforming many larger open-source and commercial models in various tasks:

VibeThinker-1.5B was compared against both reasoning-centric and non-reasoning models and consistently outperformed non-reasoning models across structured reasoning benchmarks, regardless of size.

• On AIME24 (math), it surpassed Kimi K2 (1.09T) by more than 10 points (80.3 vs. 69.6).
• On LiveCodeBench v6, it exceeded Claude Opus 4 (51.1 vs. 47.4).
• On GPQA, it performed better than GPT-4.1 and Claude, doubling its base model’s score (from 16.4 to 46.7).

This supports the argument that size is not the sole determinant of reasoning capability, and with the right training approach, smaller models can achieve or even surpass the performance of significantly larger systems in targeted tasks.

While VibeThinker excels in structured logical tasks, it may lag behind in general knowledge reasoning compared to larger models, highlighting a potential trade-off between specialization and capacity.

Implications for Enterprise Adoption

The release of VibeThinker-1.5B includes recommended inference settings and highlights its suitability for deployment on edge devices, offering cost-efficient reasoning systems for various applications.

This positions VibeThinker-1.5B not only as a research breakthrough but also as a practical solution for enterprises seeking locally deployable reasoning systems at reduced costs.

Weibo’s Strategic Position

Weibo, a major player in China’s social media landscape, is adapting to market challenges by exploring new avenues such as AI research and development, as evidenced by the launch of VibeThinker-1.5B. This shift in focus demonstrates Weibo’s ambition to expand its technological prowess beyond traditional social media operations.

Implications for Enterprise Decision Makers

For engineering leaders and AI teams in enterprises, VibeThinker-1.5B offers practical insights into optimizing pipelines, cost-effectiveness, and deploying reasoning-capable models within existing systems. Its innovative training methodology and benchmark transparency provide a roadmap for enhancing smaller models’ performance without relying solely on large-scale pretraining.

VibeThinker-1.5B represents a new class of compact, reasoning-optimized models that present viable solutions for enterprise applications, offering a balance of cost-efficiency, latency reduction, interpretability, and control.

Ultimately, VibeThinker-1.5B signifies a significant advancement in the field of AI research, with implications for practical enterprise implementations and the broader landscape of Chinese open-source AI offerings.