Weibo's VibeThinker-3B AIME 2026 score rests on a post-training pipeline
VibeThinker-3B, a 3-billion-parameter language model from a nine-person team at Sina Weibo, posts benchmark scores that put it alongside models more than 200 times its size. On AIME 2026 the paper reports 94.3, ahead of Gemini 3 Pro's 91.7 and within reach of DeepSeek V3.2, a 671-billion-parameter system. The scores stand on the paper's reported methodology, but the technical report's substance is the post-training pipeline that produced them, and the explicit hypothesis the team builds on top: that verifiable reasoning compresses into small models in a way broad knowledge does not.
The team's pipeline, which they call the Spectrum-to-Signal Principle, is the technical core of the paper. VibeThinker-3B is post-trained on Qwen2.5-Coder-3B, a compact foundation model from Alibaba's Qwen team, across four sequential phases. A two-stage supervised fine-tuning process applies curriculum learning: the first pass covers a broad mix of math, code, STEM, dialogue, and instruction data, and the second narrows to harder, longer-horizon reasoning problems. The team also discards reasoning traces shorter than 5,000 tokens and removes problems the earlier VibeThinker-1.5B could solve more than 75 percent of the time, leaving the model to focus on genuinely hard cases. These curation choices are explicit and reproducible; that alone makes the work more than a leaderboard entry.
Reinforcement learning follows under a custom algorithm called MaxEnt-Guided Policy Optimization, or MGPO, that prioritizes problems at the model's capability boundary rather than solved or unsolvable items. The paper reports a counterintuitive finding: a progressive context-window expansion strategy that worked at 1.5B parameters actively hurt performance at 3B. The team hypothesizes the difference is the stronger starting checkpoint: truncating reasoning traces during warm-up stopped removing noise and started disrupting valid reasoning. The fix was a single 64,000-token context window across training. A secondary optimization, called Long2Short Math RL, redistributes reward to favor shorter correct solutions, compressing the model's reasoning style without an accuracy penalty. The cumulative effect of these design decisions is what produces the AIME 2026 score, not the choice of base model.
The third phase distills high-quality reasoning trajectories from the RL checkpoints back into a unified model. The team ranks candidate traces by a learning-potential score, the student model's perplexity on each teacher trajectory, prioritizing correct solutions the student has not yet internalized. The final phase, Instruct RL, applies reinforcement learning on instruction-following tasks using rule-based validators for format constraints and rubric-based reward models for open-ended quality. None of this is novel in isolation; the contribution is the assembly. Researchers reproducing the recipe can test whether the same scaling result holds for other base models, which is the real test of whether the result is architectural or curatorial.
Wrapped around the pipeline is the team's Parametric Compression-Coverage Hypothesis, which separates capability into two regimes. Verifiable reasoning, where answers can be checked, is a parameter-dense capability that compresses into a small core. Open-domain knowledge is parameter-expansive, spreading across many more parameters because the data itself is broad. The paper tests this claim directly: on GPQA-Diamond, a graduate-level science benchmark, VibeThinker-3B scores 70.2, well behind Gemini 3 Pro's 91.9 and Claude Opus 4.5's 87.0. The authors treat that gap as confirmation rather than contradiction, and the framing is what gives the work its intellectual weight. The hypothesis predicts the gap, the model produces it, and the consistency is what carries the argument forward.
The most direct attack on the benchmark results is benchmark contamination, and the community response captured this clearly. The paper's strongest evidence on this front is the LeetCode block, with contests dated April 25 to May 31, 2026, dates that postdate any plausible training cutoff. Under identical evaluation conditions, VibeThinker-3B passed 123 of 128 first-attempt submissions, beating GPT-5.2, Doubao Seed 2.0 Pro, Kimi K2.5, and Claude Opus 4.6. AIME 2025 and 2026 are different propositions; competition math problems have circulated widely, and the decontamination process the paper describes, an n-gram-based filter, is, as critics in the source note, a thin guard against this class of leakage. The LeetCode result is harder to dismiss, but it is one block of evidence, not a structural defense.
The gap between benchmark performance and practical utility is where the community response turns harsh. Users running the model in LM Studio reported that follow-up questions regressed to the first prompt, a failure pattern the source's benchmarks do not measure. Others noted that the model did not recognize uv scripts, a basic Python tooling concept. These are anecdotal reports, not controlled experiments, but they suggest the post-training pipeline optimized for closed-form reasoning at the expense of the conversational robustness that production assistants require. The paper does not claim production readiness, and the authors scope the contribution to verifiable reasoning, but the gap is large enough that a small-model reasoning engine will not displace a large generalist on its own.
The cost framing is what gives the result its industry weight. The team reports the post-training run for the 1.5B predecessor cost $7,800, compared to the $294,000 estimated for DeepSeek R1. The source does not report the 3B version's post-training cost, but the order-of-magnitude claim bears on the economics of post-training small reasoning models at constrained budgets. The scaling hypothesis is not being overturned; it is being carved up into regimes where size matters and regimes where it does not.
The pipeline generalizes or it does not, and the paper does not test that question. It runs VibeThinker-3B on top of Qwen2.5-Coder-3B without reporting parallel results on another small foundation, and it does not extend the verifiable-reasoning evaluation beyond math and code. The hypothesis predicts broader compression, but the prediction is unverified. Until those experiments run, the work stands as a proof that small reasoning models can match much larger ones on the narrow slice the paper targets, and as a framework for why the rest of the capability stack is harder to compress.