Vibethinker-1.5B, Vibethinker’s new open-source AI model, improved DeepSoke-R1 on a budget of $7,800 after training.

Another day at the end of 2025, another impressive result from a Chinese company in open source artificial intelligence.

Chinese social networking company Vibe’s AI division recently released its open source Vibethinker-1.5b1.5 is a 1.5 billion parameter large language model (LLM) that is an excellent variation of a rival Chinese tech firm. Alibaba’s QWEN2.5-MATH-1.5B.

It is now available for free download and use by researchers and enterprise developers – even for commercial purposes – under a permissive MIT license. Hug facefor , for , for , . GitHub And Modelsscopewith a Technical report At the open access science publishing site arxiv.org.

And yet, despite its compact size, the Webtanker-1.5B achieved benchmark-topping performance on math and code tasks, rivaling or exceeding models hundreds of times its size, even besting Chinese rival Dapsek’s popular R-One earlier this year.

It further eclipses AI’s magisterial medium and holds its own against Anthropic’s Cloud Ops 4 and Openai’s GPT-OSS-20B medium, while requiring a fraction of the infrastructure and investment.

It is also the case that post-training is achieved on a budget of just US$7800 for computational resources (3900 GPU hours on NVIDIA H800s)—less than the tens, or even hundreds of thousands of dollars typically required to fine-tune similar or larger-scale models.

Note that this is not the total cost of model development, however: LLM is trained in stages. The first is done before training, when the model learns basic language structure and general knowledge by predicting the next word in large amounts of text from the Internet, books, and articles. This gives it fluency but not much sense of how to follow instructions or communicate

After training, using very small, high-quality datasets—for example, example questions, prompts, and expert-written responses—teach the model how to answer helpfully, reason through problems, and adapt to human expectations. Still, the cost-effectiveness of Weibo’s post-training training on Webtanker-1.5B is remarkable and should be appreciated.

Open source parameters release assumptions about scale, compute intensity, and minimum feasible size for high-performance LLMs.

A Different Approach to Training: Spectrum to Signal

Vibethinker-1.5B is not a measure of its performance, but of the training framework behind it: the spectrum-to-signal principle (SSP).

Instead of fully optimizing a model for single-response accuracy (pass@1), the SSP framework decols supervised fine-tuning (SFT) and reinforcement learning (RL) in two separate phases with different goals.

• SFT (“Spectrum Phase”): The model is trained to maximize the diversity of possible correct answers, thereby improving its pass score. It produces a wide range of respectable solution paths.

• RL (“Signal Phase”): The second step uses a reinforcement learning system (called Maxent Guided Policy Optimization, or MGPO) to identify and optimize the most correct paths from this diverse solution pool. MGPO prioritizes problems where the model is most uncertain, using entropy-based weights to focus on learning.

The authors argue that this separation allows smaller models to explore the reasoning space more efficiently.

Vibethinker-1.5B makes a compelling case that the industry’s reliance on parameter scaling as the only path to better reasoning performance may be obsolete.

By adopting a diversity-first training pipeline, Webui has demonstrated that smaller, more accessible models can match and even outperform multibillion-dollar systems in logic-heavy tasks.

The low resource footprint is one of the most important aspects of Vibethinker-1.5b. At under $8,000, the post-training cost is 30–60x less than models like the Deepsec R1 and MINIMAX-M1, which cost between $294K and $535K to train.

Performance across domains

Despite its small size, Vibethinker-1.5B provides cross-domain reasoning that outperforms many large open source and commercial models.

Webethinker was benchmarked against both argument-centric models (Magistral, Claude, Openei O3-Mini) and non-argumentative LLMs (GPT-4.1, Kimi K2, DepSeek V3). In the Structured Reasoning benchmark, the model consistently outperformed the non-reasoning models, regardless of:

• On Aime24 (Math), he beat Kimi K2 (1.09T) by 10 points (80.3 vs 69.6).

• On LiveCodebench V6, it outperformed Cloud Ops 4 (51.1 vs 47.4).

• On GPQA, it scored below GPT-4.1 and Cloud, but still doubled its base model (from 16.4 to 46.7).

This supports the authors’ claim that size is not the only criterion for capacity—with proper training design, small models can reach or even exceed the performance of much larger systems in targeted tasks.

In particular, it achieves parity with models hundreds of times larger on math and code, although it lags behind in general cognitive reasoning (GPQA), where larger models retain an edge.

This suggests a potential specialization trade-off: while Webinar excels at structural logic tasks, it lacks the capacity for extensive encyclopedic recall, a well-known limitation of miniature architectures.

Guidance for Enterprise Adoption

The release includes recommended estimation settings (temp = 0.6, top_p = 0.95, maxtokens = 40960).

This model is small enough to be deployed on edge devices, including mobile phones and in-vehicle systems, while costing an estimated 20-70x cheaper than larger models.

This positions Webtanker-1.5B not just as a research achievement, but as a potential basis for a cost-effective, spatially deterministic reasoning system.

Weibo’s strategy and market position

Weibo, launched in 2009 by Sina Corporation, remains the cornerstone of China’s social media ecosystem. Often described as China’s version of X (formerly Twitter), the platform combines microblogging, multimedia content, and trending topic features with a regulatory environment shaped by strict government oversight.

Despite counting 600 million monthly active users (more than double X), Investors are not optimistic about its potential to grow advertising revenue In the near future, and Weibo is facing fierce competition from video-first platforms like Dwayne, which are attracting younger users and increasing time elsewhere.

In response, Weibo has leaned into creative monetization, live streaming, and vertical video.

The platform’s role as a digital public square also makes it the focus of regulatory scrutiny. Chinese authorities continue to apply pressure on issues ranging from content governance to data security. In September 2025, Weibo was among the platforms cited in the official warningshighlighting its ongoing exposure to policy risks.

Vibe’s push into AI R&D is exemplified by the release of Vibethinker-1.5b. In addition to being a media platform, Weibo is positioning itself as a player in the next phase of Chinese AI development, using its capital reserves, user behavior data, and in-house research capacity to tap into adjacent technological domains.

What does this mean for enterprise technology decision makers?

For engineering leaders and enterprise AI teams, Webtanker’s release has practical implications for everything from orchestration pipelines to cost modeling.

A 1.5b-parameter model that outperforms larger models by 100x on math and programming tasks doesn’t just save compute—it changes the architectural balance. This enables LLMs to be deployed on limited infrastructure, reduces latency at the edge, and lowers the barrier to entry for applications that otherwise require API access to closed, frontier-scale models.

This is of value to enterprise MLs that seek to deploy reason-capable agents into existing systems, or for platform owners tasked with integrating MLs into automated workflows.

It also addresses those who manage learning from human feedback (RLHF) pipelines for reinforcement learning or individual optimization in hybrid cloud environments.

The model’s post-training methodology—specifically its entropy-targeted reinforcement learning approach—provides a roadmap for teams looking to optimize small checkpoints instead of relying on large-scale parameterization.

Vibethinker’s benchmark transparency and data elimination initiatives also highlight another emerging priority in enterprise AI: auditability. Although its performance on general knowledge tests still trails larger frontier models, its task-specific reliability makes it an attractive candidate for controlled environments where accuracy is more important than coverage.

In short. , Vibethinker-1.5b is not just a research milestone—it’s a strong candidate for practical business use, deployment, and learning. This suggests that a new class of compact, reasoning-optimized models is viable for enterprise use cases that were previously the domain of much larger systems. For organizations trying to balance cost, latency, interpretation and control, it’s a nice new option to the long, growing list of Chinese open source offerings.