VibeThinker-3B: Weibo's 3-Billion-Parameter Model Scores 94.3 on AIME 2026
VibeThinker-3B is a 3-billion-parameter reasoning model from Weibo (Sina Weibo) that scores 94.3 on the AIME 2026 math benchmark. Released in June 2026 as the arXiv technical report 2606.16140, the paper argues that a "Spectrum-to-Signal" post-training pipeline can elicit large-model verifiable reasoning from a small model. The authors report that the same model matches or exceeds flagship models hundreds of times larger, such as DeepSeek V3.2, GLM-5, and Gemini 3 Pro, on math and coding reasoning.
Spread the Spectrum First, Then Narrow to the Signal
The design axis of VibeThinker-3B is training order. The "Spectrum-to-Signal" pipeline runs in three stages. The first, curriculum-based supervised fine-tuning (SFT), raises difficulty gradually while securing broad solution diversity up front. Next, multi-domain reinforcement learning (RL) across math, coding, and other domains amplifies the verifiable correct-answer signal among the candidate solutions spread out earlier. The final stage, offline self-distillation, compresses the strengthened ability back into the model.
The authors frame this as a "Parametric Compression-Coverage Hypothesis." The core idea is that small models should not narrow straight to correct answers but should widen coverage first and then converge on the signal. Read another way, it is an attempt to route around a small model's tight parameter budget through the order in which it learns.
From AIME 94.3 to LeetCode 96.1%: How to Read the Numbers
VibeThinker-3B's results cluster in gradable domains. The main reported scores are as follows.
- AIME 2026 math: 94.3 (97.1 with claim-level test-time scaling).
- AIME 2025: 91.4, and HMMT 2025: 89.3.
- LiveCodeBench v6: 80.2 Pass@1.
- Unseen LeetCode weekly and biweekly contests from late April to late May 2026: 96.1% acceptance rate.
On IFEval, which measures instruction following, VibeThinker-3B scored 93.4. The line to watch is number four. A 96.1% acceptance rate on "unseen" contests held after the training window can be read as evidence that sidesteps concerns about benchmark contamination. That unseen coding result arguably carries more weight than the competition-math scores as a sign of real generalization.
How Far Does the "On Par With Giants" Claim Actually Hold
The paper's abstract states that the model matches or exceeds flagship models hundreds of times larger, such as DeepSeek V3.2, GLM-5, and Gemini 3 Pro. Some secondary outlets stretched the comparison to include Claude Opus 4.5, running headlines like "a 3B model that beats Opus."
This comparison reignited a benchmark debate. VentureBeat relayed criticism that a small model's benchmark edge is confined to narrow math and coding tasks like AIME and LiveCodeBench, and that heavy post-training optimization on those same tasks can widen the gap with general capability. In short, "beating the giants" reads accurately as a result on specific reasoning benchmarks, not parity across all domains.
A Crack in the Parameter-Performance Curve
VibeThinker-3B's significance lies in the direction it points, not a single top score. As of 2026, reasoning performance was largely assumed to scale with parameter count and training budget, so a 3-billion-parameter model reaching 94.3 on AIME cracks that assumption. The practical implication is that post-training built on verifiable rewards can sharply raise a small model's reasoning.
But the causality should be read carefully. The result is less "3B is enough" and more a conditional claim: "for narrow tasks whose answers can be graded, 3B can be pushed surprisingly far." That the reported strengths concentrate in math and code follows directly from this condition.
How a Practitioner Should Actually Use This
The fact that a small open-weight model can go toe-to-toe with far larger ones in verifiable domains has immediate value for on-premise and edge deployment. For tasks where the reward is clearly defined, such as internal math and code grading, automated scoring pipelines, or tool calls whose outputs can be verified, a 3-billion-class model leaves real room to cut cost.
The reverse also holds: generalization to open-ended dialogue, knowledge, and safety needs further verification. The safest way to weigh adoption is to ask first whether your task's answers can be graded automatically. The more ambiguous the grading criterion, the less these benchmark scores transfer intact.
AI & tech,
read in depth
Beyond the headlines — into the context and the structure
AGI Soon As Possible · asapai.co.kr