Weibo AI open-sources a $7,800 reasoning model

VibeThinker-1.5B is an open-source, 1.5-billion-parameter reasoning model from Weibo AI built to tackle hard math and competitive coding puzzles on a shoestring budget.

Trained on a total budget of just $7,800, this compact dense model is built on top of Qwen2.5-1.5B and delivers reasoning accuracy that punches way above its weight class.

What it can do

Weibo AI designed VibeThinker-1.5B specifically for competitive-style math and algorithmic coding, such as problems from LeetCode or Codeforces.

Its benchmark numbers are genuinely startling for its size. Across three major competition-level math benchmarks, AIME24, AIME25, and HMMT25, it posted scores of 80.3, 74.4, and 50.4.

Those numbers outperform the original DeepSeek R1, a model with more than 400 times as many parameters that scored 79.8, 70.0, and 41.7 on the same tests. Weibo AI notes that its overall reasoning performance is comparable to larger systems like GPT OSS-20B Medium.

On code generation, it scored 55.9 on LiveCodeBench v5 and 51.1 on v6, edging out Magistral Medium's 50.3.

The catch is that this is strictly an experimental release. Weibo explicitly advises against using it for general-purpose chat or day-to-day writing, and it works significantly better when prompts are written in English.

How it works and how to run it

The magic behind squeezing this much logic into a tiny budget comes down to what Weibo calls the Spectrum-to-Signal Principle (SSP).

During supervised fine-tuning, the team prioritizes generating a wide diversity of solution paths. Once that broad problem-solving base is established, reinforcement learning steps in to reinforce the correct logical signals.

Getting it running is practically effortless. The model is released under an MIT license, with weights hosted on Hugging Face and ModelScope alongside an arXiv technical report.

Because it only has 1.5 billion parameters, you can run it locally on almost any consumer GPU. It integrates directly with Transformers, vLLM, and SGLang, and Docker images are ready to go.

Weibo recommends setting the max generation length to 40,960 tokens with a temperature of 0.6 or 1.0 so the model has enough runway to think through long reasoning steps.

Why it matters

If you build local coding assistants or practice competitive programming, VibeThinker-1.5B is well worth downloading. Getting frontier-style math reasoning on hardware that fits inside a laptop without cloud API costs is a massive practical win.

It also shows that test-time compute and smarter RL pipelines can break the brute-force scaling barrier. You will not want to use it to draft emails, but as a dedicated logic engine for algorithmic tasks, it proves tiny models can punch far above their weight.