Algorithm Community (SuanNi) is an open-source ecosystem community integrating "AI Large Model Development Services + Algorithms + Compute Power" — welcome to follow!
Can a 3-billion-parameter model challenge hundred-billion, even trillion-parameter flagship giants in math competitions and real-world coding?
Weibo's newly open-sourced VibeThinker-3B pushes small-model performance in specific capability dimensions to the limit.
It scores 94.3 on AIME26, and with test-time scaling via Claim-Level Reliability Assessment (CLR) pushes to 97.1. On recent LeetCode Weekly Contests, its first-submission pass rate hits 96.1%.
A 3B small model takes verifiable reasoning to near-ceiling performance, matching top-tier open-source hundred-billion and trillion-parameter models — even surpassing closed-source Claude Opus 4.5.
3B Model Posts Flagship-Grade Scores
VibeThinker-3B is the latest exploration in the VibeThinker series at the 3-billion-parameter scale; its predecessor VibeThinker-1.5B already validated the approach at 1.5B parameters.
This scale-up focuses on math, coding, and STEM — reasoning tasks with clear verification signals.
The model follows the Spectrum-to-Signal (S2S) post-training paradigm: curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation drive verifiable reasoning to the extreme.
Math competition results are the most striking. On AIME26 it scores 94.3 out of the box; combined with Claim-Level Reliability Assessment (CLR), a test-time scaling strategy for verifiable reasoning, it reaches 97.1. On IMO-AnswerBench — a high-difficulty benchmark of 400 IMO-level problems — the single-pass score is 76.4, rising to 80.6 with CLR. On traditional benchmarks like HMMT and AIME, performance remains consistently strong.
For comparison, DeepSeek V3.2 with 671B parameters scores 78.3 on IMO-AnswerBench, GLM-5 with 744B scores 82.5, and Kimi K2.5 with 1 trillion scores 81.8. VibeThinker-3B's 80.6 squeezes into that bracket — at a fraction of their parameter counts.
Beyond math, VibeThinker-3B excels across coding, knowledge, and instruction following.
On LiveCodeBench v6 it achieves 80.2 Pass@1; on IFEval it scores 93.4.
Across multiple verifiable-reasoning benchmarks, VibeThinker-3B competes head-to-head with first-tier reasoning models like Qwen3.6 Plus, Gemini 3 Pro, GLM-5, and Kimi K2.5, landing squarely in the top-model performance band.
Out-of-distribution generalization is another highlight. The team evaluated on LeetCode Weekly and Biweekly Contests from April 25 to May 31, 2026 — contests the model had never seen — using Python exclusively.
It passed 123 out of 124 problems on first submission — a 96.1% pass rate. The model isn't just regurgitating familiar problem sets; it solves unseen real contest problems reliably.
Code capability transfer often tells us more than math benchmarks, because problem statements, constraints, and data structures are all novel — the model must genuinely understand the task to solve it.
Layered Training Pipeline
VibeThinker-3B inherits the Spectrum-to-Signal Principle (SSP) introduced with VibeThinker-1.5B.
The core idea: supervised fine-tuning (SFT) builds a broad spectrum of plausible reasoning trajectories; reinforcement learning (RL) then amplifies the correct signal using verifiable rewards.
First comes a two-stage curriculum-based SFT. The first half covers math, code, STEM reasoning, general chat, and instruction following; the second half shifts to harder samples with longer reasoning horizons.
Diversity-Exploring Distillation is used here to preserve multiple valid solution paths, preventing the model from collapsing to a single heuristic.
Next is multi-domain reasoning RL, reusing MaxEnt-Guided Policy Optimization (MGPO). RL is applied sequentially to math, code, and STEM tasks, all within a single 64K context window, preserving full long-horizon reasoning traces.
Then comes offline self-distillation. High-quality reasoning traces from math, code, and STEM RL checkpoints are distilled back into a unified student model. Selection uses a learning-potential score, prioritizing traces where the answer is correct but the student model hasn't yet mastered the reasoning.
Finally, instruction RL. After pushing reasoning to its limits, this step improves controllability on user prompts. Data splits into format-sensitive instructions and open-ended instructions, using rule-based verifiers and rubric-based reward models respectively. The model learns to reason and to follow instructions — IFEval 93.4 is the direct result.
The Parameter Compression-Coverage Boundary
From VibeThinker-1.5B to VibeThinker-3B, the team's goal has been to probe the true boundary of small models along specific capability axes.
Observing these results, they propose the Parametric Compression-Coverage Hypothesis. The core claim: different capabilities depend on parameter scale in fundamentally different ways.
Verifiable reasoning behaves like a highly compressible, parameter-dense capability. Its essence is a loop of multi-step reasoning, constraint satisfaction, self-correction, and answer verification. The task space has clean structure and reliable feedback signals — giving small models a shot at frontier performance.
A math answer is either right or wrong; code either runs or it doesn't. Verification signals are clean. Parameters here act more like amplifiers of latent potential than as storage for the world.
Open-domain knowledge, general chat, and long-tail scene understanding are a different story. These rely on massive parameters to broadly cover facts, concepts, and world knowledge. An unseen fact can't be answered; an untrained long-tail scenario can't be handled. Compression doesn't work. Parameters here serve storage and retrieval — without scale, there's no coverage.
If the hypothesis holds, the small-vs-large model relationship isn't replacement but complementarity. In domains with clear feedback mechanisms, small language models (SLMs) form an independent frontier path. Compression-type and coverage-type capabilities may need different scaling strategies; forcing a single parameter-scaling logic onto both leaves both underserved.
VibeThinker-3B draws a different capability curve. For reasoning — structured tasks with reliable feedback — even aggressive parameter compression to 3B can reach the 1T-flagship performance band. For knowledge, chat, and long-tail understanding — capabilities demanding broad coverage — parameters remain unavoidable.
Pursuing this path, frontier performance may no longer depend solely on parameter stacking. In verifiable domains, pushing reasoning to its limits makes small models a viable independent pole.
Weibo's open-sourcing of VibeThinker-3B validates this path's feasibility. Whether the community can take it a step further is well worth watching.
References:
https://huggingface.co/WeiboAI/VibeThinker-3B
https://github.com/WeiboAI/VibeThinker