A nine-person team at a Chinese social media company just embarrassed the biggest names in AI with a model that's hundreds of times smaller than the competition
A language model with 3 billion parameters just matched the reasoning performance of systems that are 200 times its size. The team behind it doesn’t work at OpenAI, Google DeepMind, or Anthropic. They work at a microblogging company.
Sina Weibo, the Chinese social media platform most people associate with viral posts rather than frontier AI research, published a 14-page technical report on arXiv detailing VibeThinker-3B. The model scored 94.3 on AIME 2026, one of the most demanding standardized math competitions in the world, placing it alongside DeepSeek V3.2 and its 671 billion parameters.
Small model, big numbers #
The benchmark results tell the story. On AIME 2026, VibeThinker-3B hit 94.3, a score that climbs to 97.1 when using claim-level test-time scaling. On LiveCodeBench v6, a coding benchmark, it posted a Pass@1 score of 80.2. The model also demonstrated superior out-of-distribution performance on recent LeetCode contests, often matching or beating those much larger systems.
The model is built on top of Qwen2.5-Coder-3B as its base architecture. The Sina Weibo team, comprising nine researchers including Sen Xu, Shixi Liu, and Wei Wang, enhanced performance through a combination of curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation techniques. The paper also introduces the Parametric Compression-Coverage Hypothesis, which offers a theoretical framework for why smaller models can punch above their weight in structured reasoning tasks.
The efficiency arms race #
VibeThinker-3B didn’t come out of nowhere. Its predecessor, VibeThinker-1.5B, launched in November 2025 and used the Spectrum-to-Signal Principle to achieve impressive results at a training cost of roughly $7,800. For context, training frontier models at companies like OpenAI and Google typically costs tens or hundreds of millions of dollars.
The 3B version extends the approach from the earlier model with more sophisticated training techniques, including curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation.
Both the model weights and codebase are fully open. Weights are available on Hugging Face and code is hosted on GitHub, both under the MIT license.
What this means for the AI investment landscape #
VibeThinker-3B has no direct connection to crypto, blockchain, or tokenization. The researchers didn’t mention anything remotely related to Web3.
For the decentralized AI projects in the crypto space that are working on inference, model hosting, and compute marketplaces, the efficiency trend is constructive. Smaller, high-performance models are easier to run on distributed networks, require less specialized hardware, and are more practical for edge deployment. A 3-billion-parameter model that performs like a 671-billion-parameter model is exactly the kind of development that makes decentralized inference economically viable rather than a theoretical exercise. When models this capable are released freely under an MIT license, the barrier to building competitive AI applications drops dramatically. That is potentially challenging for any company, crypto or otherwise, that is trying to monetize model access as a core revenue stream.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our