Weibo has launched VibeThinker-1.5B, an open source LLM that delivers frontier-level reasoning at a fraction of usual costs.
Weibo has released VibeThinker-1.5B, a fully open source 1.5-billion-parameter language model fine-tuned from Alibaba’s Qwen2.5-Math-1.5B. Published under the permissive MIT Licence, the model is available for free download and commercial integration on Hugging Face, GitHub, and ModelScope, with its technical report openly accessible on arxiv.org.
Despite its compact scale, VibeThinker-1.5B delivers benchmark-leading reasoning performance, surpassing DeepSeek R1 (671B) on formal reasoning tasks and matching models such as Magistral Medium, Claude Opus 4, and OpenAI gpt-oss-20B Medium. It posts strong cross-domain scores, including 74.4 on AIME25, 51.1 on LiveCodeBench v6, and 46.7 on GPQA-Diamond. It also beats Kimi K2 (1.09T) on AIME24 and significantly raises GPQA performance compared to its base model.
One of its defining achievements is cost efficiency: the model’s post-training required only $7,800 USD, built on 3900 GPU hours using Nvidia H800s—30–60× cheaper than comparable large-scale systems such as DeepSeek R1 and MiniMax-M1.
VibeThinker-1.5B is powered by the Spectrum-to-Signal Principle (SSP), which splits supervised fine-tuning and reinforcement learning into diversity generation and signal amplification phases. This approach enables small models to achieve competitive reasoning capability without relying on massive parameter counts.
Enterprise teams stand to benefit from its low inference cost, edge-deployable footprint, and recommended settings that keep latency and compute overhead minimal. Its transparency, data decontamination, and audit-friendly benchmarks further strengthen its appeal for controlled, cost-sensitive environments.
The release marks Weibo’s deepening shift into AI R&D and positions China’s open-source ecosystem as a formidable competitor in high-reasoning LLM innovation.
