ウェイボーの小型モデル「VibeThinker-3B」がベンチマーク論争を再燃させた理由（15 分読了）

On Sunday, a team of nine researchers at Sina Weibo — the Chinese social media giant better known for its microblogging platform than for cutting-edge artificial intelligence — quietly posted a 14-page technical report to arXiv that sent shockwaves through the AI research community. Their claim: a language model with just 3 billion parameters can match or exceed the reasoning performance of flagship systems from Google DeepMind, OpenAI, Anthropic, and DeepSeek that are hundreds of times larger.

The model, called VibeThinker-3B, scored 94.3 on AIME 2026 — the American Invitational Mathematics Examination, one of the most demanding standardized math competitions in the world. That figure places it alongside DeepSeek V3.2, a model with 671 billion parameters, and ahead of Gemini 3 Pro, Google's high-performance flagship reasoning system, which scored 91.7. With a test-time scaling technique the team calls Claim-Level Reliability Assessment, the score climbs to 97.1, edging past virtually every system in the public record.

Within hours of publication, the paper had drawn 62 upvotes on Hugging Face's daily papers feed, the model repository had accumulated 130 likes, and the GitHub repository had reached 685 stars. But the reaction on social media was not uniformly celebratory. It was, in many cases, deeply skeptical.

"WHAT THE HELL is happening in AI?" wrote the user @orcus108 on X, in a post that accumulated over 161,000 views. "A 3B parameter model just put up coding benchmark scores in the same league as Claude Opus 4.5… I genuinely don't know if this is a breakthrough or if the benchmarks are broken."

That tension — between genuine scientific advancement and the growing suspicion that AI benchmarks have become gameable to the point of meaninglessness — sits at the heart of the VibeThinker-3B story. And the answer matters enormously, not just for academic bragging rights, but for the multibillion-dollar question of whether the AI industry's relentless push toward ever-larger models is the only path to intelligence.

Benchmark scores that defy the scaling laws of modern AI

The results reported in the technical report are, by any conventional standard, extraordinary.

On the mathematics side, VibeThinker-3B achieved 91.4 on AIME 2025, 94.3 on AIME 2026, 89.3 on HMMT 2025 (the Harvard-MIT Mathematics Tournament), 93.8 on BruMO 2025 (the Brown University Math Olympiad), and 76.4 on IMO-AnswerBench, a benchmark comprising 400 problems at the level of the International Mathematical Olympiad. In coding, it posted an 80.2 Pass@1 on LiveCodeBench v6, a benchmark designed to test executable code generation, and achieved a 96.1 percent acceptance rate on unseen LeetCode weekly and biweekly contests from late April through late May 2026. On instruction following, it scored 93.4 on IFEval.

To put the parameter disparity in perspective: DeepSeek V3.2 has 671 billion parameters — roughly 224 times the size of VibeThinker-3B. GLM-5, from Zhipu AI, has 744 billion parameters. Kimi K2.5, from Moonshot AI, exceeds 1 trillion. VibeThinker-3B's 3 billion parameters could run on a consumer laptop.

The researchers frame this result not as an anomaly but as evidence for a broader theoretical claim. They introduce what they call the "Parametric Compression-Coverage Hypothesis," which argues that different types of AI capability have fundamentally different relationships to model size. Verifiable reasoning — the kind tested by math competitions and coding challenges, where answers can be definitively checked — is what the paper calls a "parameter-dense" capability: one that can be compressed into a compact core. Open-domain knowledge, by contrast, is "parameter-expansive," requiring broad coverage across facts, concepts, and edge cases that inherently demands more parameters.

The paper acknowledges this distinction directly. On GPQA-Diamond, a graduate-level science knowledge benchmark, VibeThinker-3B scored just 70.2 — well behind the 91.9 achieved by Gemini 3 Pro and the 87.0 scored by Claude Opus 4.5. The authors write that this gap "is consistent with our claim rather than a contradiction to it: the main finding is not that a 3B model has fully replaced leading general-purpose models, but that a small model can reach first-tier performance on many verifiable reasoning tasks."

Inside the four-stage training pipeline that powers a tiny reasoning engine

VibeThinker-3B is not built from scratch. It is post-trained on top of Qwen2.5-Coder-3B, a compact foundation model from Alibaba's Qwen team, through what the Weibo AI researchers call the "Spectrum-to-Signal Principle" — a multi-stage pipeline first introduced in the team's earlier VibeThinker-1.5B work in November 2025.

The training unfolds in four major phases. The first is a two-stage supervised fine-tuning process that uses curriculum learning: the model first trains on a broad mixture of math, code, STEM reasoning, general dialogue, and instruction-following data, then shifts to a curated subset of harder, longer-horizon reasoning problems. In the second stage, samples with reasoning traces shorter than 5,000 tokens are discarded, and problems that VibeThinker-1.5B can solve more than 75 percent of the time are filtered out, forcing the model to focus on genuinely difficult challenges.

The second phase applies reinforcement learning across multiple domains — mathematics, code, and STEM — using the team's MaxEnt-Guided Policy Optimization algorithm, or MGPO, which prioritizes training on problems at the model's current capability boundary rather than problems it already solves easily or finds impossible. Notably, the team found that a strategy that worked well at the 1.5B scale — progressively expanding the context window during RL training — actually hurt performance at 3B. They hypothesize that the stronger starting checkpoint meant that truncating reasoning traces during warm-up was no longer removing noise but disrupting valid reasoning patterns. The solution was to train with a single 64,000-token context window throughout.

Within the math RL phase, the team also introduces what it calls "Long2Short Math RL," a secondary optimization stage that redistributes rewards to favor shorter correct solutions over longer ones, reducing verbosity without sacrificing accuracy. The technique uses a zero-sum reward redistribution that avoids biasing the overall reward signal while nudging the model toward more efficient reasoning.

The third phase extracts high-quality reasoning trajectories from the RL-trained checkpoints and distills them back into a unified model through supervised fine-tuning. The team uses a "learning-potential score" — essentially the student model's perplexity on each teacher trajectory — to prioritize traces that are correct but that the student has not yet internalized. The final phase, called Instruct RL, applies reinforcement learning on instruction-following tasks using a combination of rule-based validators for format constraints and rubric-based reward models for open-ended quality assessment.

Francesco Bertolotti, an AI researcher who flagged the paper early on X, described the approach succinctly: "These results were achieved primarily through post-training refinements on Qwen2.5-Coder. The paper doesn't provide many details, but it appears they distill from RL ckpts and then do a final RL-based instruct RL." His post drew over 161,000 views.

Real-world testing reveals the gap between benchmark scores and practical AI performance

For every enthusiastic reaction, the paper drew an equally forceful objection. The AI research community in mid-2026 has grown deeply wary of benchmark-driven claims, and VibeThinker-3B arrived in an environment primed for suspicion.

"The benchmarks are literal pattern matching single file coding," wrote @BigMoonKR on X. "It has no relation to actual coding work. I don't know how people still don't get this."

"Benchmaxxing," declared @oflu_bedirhan, using a term that has become shorthand in the AI community for models that appear optimized specifically for benchmark performance at the expense of real-world utility.

The most pointed criticism came from users who actually downloaded and tested the model. "Just tried the full precision," wrote @politilols. "It doesn't even know what a uv script (so the most popular Python dev tool) is. Haven't seen that in a single LLM in at least a year now. Benchmaxxed." When Bertolotti responded that the model seemed more focused on mathematical reasoning than practical coding, the user countered: "They include a livecodebench score. Zero chance that is reflective of the model."

@Itsdotdev raised a structural criticism: "Look into the benchmarks themselves and it probably won't be so shocking. Why no DeepSWE? Why none of the standard benchmarks SOTA providers use?" The user @AvenirReym posed a more diagnostic question: "If it holds on a benchmark made after the model's training cutoff, it's real. If it only wins on AIME-style sets that have been circulating for years, it's leakage."

The paper's authors appear to have anticipated these objections. The technical report states that training sets "have undergone strict benchmark decontamination," including n-gram-based filtering to remove "n-gram overlaps with evaluation sets."

The LeetCode contest evaluation — which covers contests from April 25 to May 31, 2026, dates that postdate any plausible training data cutoff — represents the most robust guard against data contamination concerns. On those contests, VibeThinker-3B passed 123 out of 128 first-attempt submissions, a 96.1 percent rate that exceeded GPT-5.2, Doubao Seed 2.0 Pro, Kimi K2.5, and Claude Opus 4.6 under identical evaluation conditions.

Still, real-world user reports suggest a significant gap between benchmark performance and practical utility — a phenomenon that has become familiar across the industry. "In LM Studio it only responds well to first question, next questions reply to the first question," reported @luismolinaab.

Why a social media company may have found a crack in the scaling hypothesis

Even the sharpest critics acknowledged that achieving these benchmark numbers at 3 billion parameters — regardless of how transferable they are to production use cases — is a meaningful engineering achievement. "Even if it's benchmaxxing doing so with 3B parameters is fascinating, goes to show how fast this field is progressing," wrote @rohityin.

The observation cuts to a question that has consumed the AI industry since the advent of the scaling hypothesis: Is bigger always better? The conventional wisdom, articulated most famously in the Chinchilla scaling laws and reinforced by the commercial dominance of ever-larger foundation models, holds that more parameters and more training data reliably yield better performance. The economic corollary is stark: training and deploying frontier models costs tens or hundreds of millions of dollars, creating enormous barriers to entry.

VibeThinker-3B challenges that consensus — but only partially. The paper is careful to draw a boundary around its claims, distinguishing between tasks with "clear verification signals" and those that require broad factual knowledge. The Parametric Compression-Coverage Hypothesis explicitly argues that small models cannot replace large ones across the board.

"The true significance of VibeThinker-3B does not lie in proving that a 3B model can replace large-scale generalists," the paper states, "but rather in providing a concrete empirical signal: the development of compact models is no longer merely a passive compromise for deployment efficiency or cost control; it emerges as a promising research trajectory that is fundamentally complementary to the traditional parameter scaling paradigm."

Perhaps the most surprising element of the work is its provenance. Sina Weibo — publicly traded on Nasdaq and Hong Kong, with a market capitalization that fluctuates in the single-digit billions — is not a company typically associated with frontier AI research. Yet the VibeThinker series is Weibo's second major open-source AI contribution in seven months.

VibeThinker-1.5B, released in November 2025, demonstrated that a model with just 1.5 billion parameters could outperform the original DeepSeek R1 on several math benchmarks — a result the team achieved for what it claimed was a post-training cost of just $7,800, compared to the $294,000 estimated for DeepSeek R1.

The research team is compact — nine authors, all listed as Sina Weibo Inc. employees. The model is released under the MIT License, one of the most permissive open-source licenses available, and the weights are freely downloadable from both Hugging Face and ModelScope. Within the first day of release, community members had already created GGUF quantizations and derivative models.

Small models, big implications, and the question the AI industry can no longer avoid

The most honest assessment of VibeThinker-3B may be that it is simultaneously less and more than what the benchmarks suggest. Less, because a model that struggles with basic knowledge of popular developer tools is unlikely to replace any production-grade coding assistant anytime soon. More, because the underlying insight — that reasoning ability and factual knowledge are partially decoupled, and that the former can be compressed far more aggressively than previously assumed — has profound implications for how the industry thinks about model design, deployment economics, and the accessibility of advanced AI capabilities.

If the Parametric Compression-Coverage Hypothesis holds, it suggests a future in which small, specialized reasoning engines operate alongside large knowledge-rich models in hybrid architectures — a vision where a 3-billion-parameter model handles the logical heavy lifting while a larger system supplies the factual grounding. Such an architecture could dramatically reduce the cost of deploying AI reasoning capabilities, potentially bringing competition-level mathematical and coding performance to devices with modest hardware.

"The interesting part is that we're starting to separate knowledge from reasoning," wrote @RealLambdaFlux on X. "A small model with strong post-training can punch way above its size on tasks with clear feedback."

@cmitsakis suggested the practical endgame: "I think small models are the future for agents because they can use tools to get the knowledge and they can run fast and cheap."

Whether that future arrives through VibeThinker-3B specifically, or through the dozens of teams now racing to reproduce and extend these results, the paper has already accomplished something that no benchmark score can fully capture.

It has forced the AI community to confront an uncomfortable possibility: that for years, the industry may have been spending billions of dollars scaling up parameters to improve a kind of intelligence that could have fit, all along, on a laptop. The weights are public. The code is open. And the most important test isn't on any leaderboard — it's whether anyone can make a model this small actually useful in the real world.