Sina Weibo AI published a 3-billion-parameter model that scores 94.3 on AIME 2026, beating Gemini 3 Pro at 91.7. The model matches DeepSeek V3.2 (671B) and GLM-5 (744B) on verifiable reasoning benchmarks. It is built on the Qwen2.5-Coder-3B stack and released under an MIT license.
TL;DR: VibeThinker-3B from Weibo AI packs 3 billion parameters yet scores 94.3 on AIME 2026, beating Gemini 3 Pro at 91.7. Built on Qwen2.5-Coder-3B, it uses a novel SFT+GRPO pipeline to match models hundreds of times its size.
What Is VibeThinker-3B and Why Does It Matter?
VibeThinker-3B is a compact, MIT-licensed reasoning model from Sina Weibo AI that achieves flagship-level performance with only 3 billion parameters. According to MarkTechPost, it scores 94.3 on AIME 2026 and 80.2 on LiveCodeBench v6. Those numbers put it in the same league as models over 200 times larger.
The model is built on Qwen2.5-Coder-3B — notably, not the newer Qwen3 line. As Sebastian Raschka highlighted on Substack, the entire architecture relies on the older Qwen2.5-Coder base. This detail matters because it demonstrates that the performance jump comes almost entirely from post-training methodology rather than from a larger or newer foundation model.
The technical report is published on arXiv as paper 2606.16140. The research community has taken notice because the gap between the base model and the final result is unusually large. Post-training is doing the heavy lifting here. Most small models plateau well below these scores.
So why does this release matter? It challenges the assumption that reasoning quality scales linearly with parameter count. If a 3B model can match DeepSeek V3.2 on math, the cost of building capable reasoning systems drops dramatically. Developers can run this model locally on consumer hardware without massive GPU clusters.
The implications extend beyond benchmarks. A 3B dense model that reasons at this level could power coding assistants, educational tools, and mathematical software on edge devices. The MIT license removes commercial restrictions. That makes adoption straightforward for startups and independent developers alike.
How Does a 3B Model Match Frontier-Level Performance?
The short answer is aggressive post-training. VibeThinker-3B starts from Qwen2.5-Coder-3B, a coding-focused base model, and applies a pipeline called Spectrum-to-Signal that combines supervised fine-tuning with Group Relative Policy Optimization. The base model already understands code structure and logic. The post-training layer refines that understanding into structured reasoning.
Raschka’s analysis on his blog notes that the reported coding and reasoning results point to strong post-training as the primary driver. The base Qwen2.5-Coder-3B model does not achieve these scores on its own. The transformation happens during the SFT and GRPO stages, where the model learns to generate step-by-step reasoning chains that arrive at correct answers.
The architecture itself is a standard dense transformer with 3 billion parameters. There is no mixture-of-experts trickery or sparse activation. Every parameter is used for every token. This stands in contrast to models like DeepSeek V3.2 (671B total parameters) or GLM-5 (744B), which use MoE architectures to activate only a fraction of their parameters during inference.
How significant is the gap? Consider the parameter ratio. DeepSeek V3.2 uses roughly 224 times more total parameters than VibeThinker-3B. On AIME 2026, the performance difference is negligible. This suggests that for verifiable reasoning tasks — math problems with definite answers — the marginal value of additional parameters diminishes rapidly once a model crosses a certain capability threshold.
The model also benefits from its coding-focused foundation. Code and mathematics share deep structural similarities: both require precise logic, sequential reasoning, and unambiguous syntax. Starting from a coder base model gives VibeThinker-3B a structural advantage on reasoning tasks that other generalist 3B models lack entirely.
What Is the Spectrum-to-Signal Post-Training Pipeline?
Spectrum-to-Signal is the post-training methodology that transforms the base Qwen2.5-Coder-3B into VibeThinker-3B. The name describes a shift from broad, unfocused capability to sharp, targeted reasoning. The pipeline combines supervised fine-tuning (SFT) with Group Relative Policy Optimization (GRPO) in a multi-stage process.
In the SFT stage, the model is trained on curated reasoning traces — high-quality examples of step-by-step problem solving. This teaches the model the format and structure of extended reasoning chains. The model learns not just what the correct answer is, but how to arrive at it through logical steps. This stage shapes the model’s output into the long chain-of-thought format that reasoning tasks require.
The GRPO stage then refines this capability through reinforcement learning. GRPO is a reinforcement learning technique that optimizes the model’s policy by comparing multiple generated responses against each other. The model generates several candidate answers for each problem. Correct answers receive positive reinforcement. Incorrect ones get penalized. Over many iterations, the model converges toward reliably producing correct reasoning chains.
What makes this pipeline novel? Traditional post-training often applies either SFT or RL, not both in a structured sequence. The Spectrum-to-Signal approach uses SFT to establish a reasoning baseline, then uses GRPO to push beyond what supervised learning alone can achieve. The result is a model that not only mimics reasoning patterns but actually optimizes for correctness on verifiable tasks.
The technical report provides extensive detail on the training configuration, data curation, and hyperparameter choices. This transparency is one reason the community has responded positively. Researchers can study the methodology and potentially apply similar techniques to other base models. The pipeline is model-agnostic in principle, though the results may vary depending on the starting foundation.
How Does VibeThinker-3B Perform on Reasoning Benchmarks?
VibeThinker-3B posts benchmark numbers that rival or exceed models an order of magnitude larger. On AIME 2026, it scores 94.3, surpassing Gemini 3 Pro at 91.7. On LiveCodeBench v6, it reaches 80.2. These are verifiable benchmarks — math competitions and coding problems with definitive correct answers, not subjective evaluations.
The AIME (American Invitational Mathematics Examination) benchmark tests advanced mathematical reasoning at the high school competition level. Scoring above 94 means the model correctly solves nearly all problems. The fact that a 3B model achieves this while Gemini 3 Pro — a frontier model from Google — scores lower raises serious questions about how parameter count relates to reasoning capability.
Here is how VibeThinker-3B compares to larger models on key benchmarks:
| Model | Parameters | AIME 2026 | LiveCodeBench v6 |
|---|---|---|---|
| VibeThinker-3B | 3B (dense) | 94.3 | 80.2 |
| Gemini 3 Pro | ~480B (MoE) | 91.7 | Not reported |
| DeepSeek V3.2 | 671B (MoE) | Comparable | Comparable |
| GLM-5 | 744B (MoE) | Comparable | Comparable |
The VentureBeat report notes that these results have reignited the debate over benchmark gaming in the AI community. Critics argue that small models can be overfit to specific benchmark formats. Supporters counter that AIME and LiveCodeBench use held-out problem sets designed to prevent exactly this kind of overfitting. The benchmarks change regularly, making memorization ineffective.
The Towards AI analysis confirms that the model runs locally and produces genuine reasoning chains. The author tested the model on novel problems outside the training distribution and observed consistent performance. This suggests the capability generalizes rather than being an artifact of benchmark-specific optimization.
Coding benchmarks tell a similar story. LiveCodeBench v6 evaluates models on real programming challenges with executable test cases. A score of 80.2 means the model writes correct solutions for four out of five problems on average. For a 3B model, this is exceptional. Most models in this size range struggle to produce syntactically correct code, let alone algorithmically optimal solutions to competitive programming problems.
Can VibeThinker-3B Actually Replace Large Coding Models?
VibeThinker-3B scored 80.2 on LiveCodeBench v6, which puts it in the same tier as models over 200 times its size (MarkTechPost, 2026). That number alone makes developers wonder if they really need a massive API-backed model for everyday programming tasks. For specific, verifiable coding challenges, the answer is leaning toward no.
But context matters. LiveCodeBench evaluates self-contained algorithmic problems, not sprawling enterprise codebases with thousands of files. A 3-billion-parameter model has limited context window capacity compared to flagship models that handle 128K or more tokens. For quick functions, unit tests, or competitive programming, VibeThinker-3B is genuinely competitive. For multi-file refactoring, it hits a wall.
The economics are compelling. Running a 3B model locally costs nothing after the initial hardware investment. Compare that to API pricing for frontier models, which can run into hundreds of dollars per month for heavy coding workloads. Small teams and individual developers stand to benefit the most here.
That said, the model excels in narrow domains. It was trained specifically on mathematical reasoning and coding tasks. Ask it to draft a marketing email or summarize a legal document, and performance drops sharply. It is a specialist, not a generalist.
So can it replace large coding models? Partially. It handles algorithmic generation and mathematical reasoning beautifully. Complex, multi-repository engineering still requires bigger architectures with larger context windows and broader knowledge bases.
What Base Model Does VibeThinker Build Upon?
VibeThinker-3B is built on Qwen2.5-Coder-3B, an older coding-focused model from Alibaba’s Qwen series — not the newer Qwen3 stack (Raschka, 2026). This detail surprised many researchers. The AI community generally assumes that newer base models automatically produce better fine-tuned results. Weibo’s choice proves otherwise.
Qwen2.5-Coder-3B was released as part of Alibaba’s specialized coding lineup. It already possessed solid programming fundamentals from pre-training on extensive code corpora. Weibo’s team recognized that this foundation, when paired with their novel training pipeline, could punch far above its weight class. The base model matters less than the post-training methodology applied to it.
Sebastian Raschka specifically highlighted this counterintuitive decision in his analysis. He noted that the technical report demonstrates how careful post-training can extract frontier-level performance from an older, smaller architecture. The Spectrum-to-Signal pipeline — combining SFT with GRPO — transformed a competent but unremarkable base model into a benchmark champion.
This finding has broader implications. Developers sitting on older models should not rush to migrate every project to the latest base. Methodical reinforcement learning and curated supervised fine-tuning can breathe new life into existing 3B architectures. The training recipe often outweighs the starting ingredients.
How Can Developers Access and Run VibeThinker-3B Locally?
VibeThinker-3B is available on Hugging Face under an MIT license, meaning developers can download, modify, and deploy it commercially without restrictions (Codersera, 2026). The model weights are publicly hosted, and integration with popular inference frameworks like vLLM, Ollama, and llama.cpp has been confirmed by community testing.
Running a 3-billion-parameter model is remarkably accessible. Here is what developers need:
- GPU memory: 8 GB VRAM is sufficient for FP16 inference; 4 GB works with INT4 quantization
- RAM: 16 GB system memory minimum for smooth operation
- Storage: Approximately 6 GB for FP16 weights, 2 GB for quantized versions
- Framework: Ollama supports one-command installation via
ollama run vibethinker - API mode: vLLM provides an OpenAI-compatible endpoint for drop-in replacement
- CPU-only: llama.cpp runs the quantized model on modern CPUs, albeit slower
- Cloud option: Hugging Face Spaces hosts a free interactive demo
- Fine-tuning: The training scripts are open source on Weibo’s GitHub repository
The MIT license is a significant advantage. Many open-weight models arrive with restrictive non-commercial clauses that block production deployment. VibeThinker-3B has no such limitation, making it viable for commercial products, internal tools, or SaaS features without legal friction.
For developers already running local models, integration is straightforward. The Qwen2.5 architecture is widely supported across inference engines. Anyone who has run Qwen, Llama, or Mistral locally will feel immediately familiar with the deployment process.
Are the Benchmark Results Too Good to Be True?
The AI community has responded to VibeThinker-3B’s results with both excitement and skepticism, with VentureBeat reporting that the model has reignited debates over benchmark gaming and evaluation methodology (VentureBeat, 2026). A 3B model matching 671B and 744B parameter systems naturally invites scrutiny.
Several factors lend credibility to the claims. First, the model is open source under MIT license. Anyone can download the weights and reproduce the evaluations independently. Second, independent testers have already confirmed key results — the Towards AI community published a reproduction showing 94.3 on AIME 2026, matching the paper’s reported score (Towards AI, 2026). Third, Sebastian Raschka, a widely respected researcher, reviewed the technical report favorably.
However, legitimate concerns remain. Benchmarks like AIME and LiveCodeBench measure narrow, verifiable tasks. They do not capture real-world software engineering complexity. A model that aces Olympiad mathematics may still struggle with ambiguous requirements or legacy system integration. Benchmark performance does not always translate to practical utility.
Additionally, the training data composition raises questions. If the evaluation datasets or their close approximations appeared in the reinforcement learning reward signals, the model may be overfitting to specific problem formats. Weibo’s paper addresses contamination checks, but the broader community continues validating.
The results are probably not fabricated. They are probably narrower in practical scope than headline numbers suggest.
What Does This Mean for the Future of Small Language Models?
VibeThinker-3B demonstrates that the 3-to-7 billion parameter range remains fiercely competitive for specialized tasks (Neurohive, 2026). The assumption that meaningful performance requires 70B+ parameters is crumbling. Weibo’s results suggest that focused training pipelines can produce models that serve specific use cases at a fraction of the compute cost.
The implications extend across multiple areas of AI development:
| Domain | Impact of Small Model Advancement |
|---|---|
| Edge computing | 3B models run on consumer hardware without cloud dependency |
| Privacy | Local inference eliminates data transmission concerns |
| Cost | Eliminates per-token API charges for high-volume tasks |
| Latency | No network round-trips; responses generate in milliseconds |
| Customization | Smaller models are easier and cheaper to fine-tune |
| Deployment | No vendor lock-in or rate limiting from API providers |
The Spectrum-to-Signal training pipeline may be the most replicable contribution. If other teams apply similar SFT-plus-GRPO methodologies to different base models, the field could see a wave of hyper-efficient specialists. A 3B math model, a 3B legal model, a 3B medical model — each trained with targeted reinforcement learning on a proven base architecture.
Enterprise adoption hinges on reliability. Companies need consistent performance across diverse tasks, not just benchmark peaks. Small models will likely complement rather than replace large ones in production environments. A routing layer could direct simple queries to a local 3B model and reserve expensive API calls for genuinely complex problems.
Frequently Asked Questions
Is VibeThinker-3B really free and open source?
Yes, the model is released under the MIT license, which permits commercial use, modification, and redistribution without fees (Codersera, 2026). The model weights on Hugging Face and the training scripts on GitHub are publicly accessible. Developers can integrate VibeThinker-3B into commercial products without licensing concerns.
What hardware do you need to run VibeThinker-3B locally?
The model requires approximately 6 GB of VRAM for FP16 inference and around 4 GB with INT4 quantization, making it runnable on consumer GPUs like the RTX 3060 (Codersera, 2026). CPU-only inference is possible through llama.cpp with quantized weights, though generation speed decreases significantly. An SSD with 6 GB of free space handles the full-precision weights comfortably.
Does VibeThinker-3B actually beat Claude Opus 4.5 in general tasks?
No, VibeThinker-3B matches Opus 4.5 specifically on verifiable reasoning and coding benchmarks like AIME 2026 and LiveCodeBench v6, where it scored 94.3 and 80.2 respectively (MarkTechPost, 2026). For general conversational tasks, creative writing, or broad knowledge queries, Claude Opus 4.5 maintains a substantial lead. The comparison applies strictly to mathematical and programming evaluations.
Why is the model called VibeThinker?
The name combines “Vibe,” referencing the informal AI community term “vibe coding” popularized in 2025, with “Thinker,” denoting the model’s chain-of-thought reasoning capabilities (Medium, 2026). Weibo’s team chose the name to reflect the model’s ability to produce code that aligns with developer intent through structured reasoning. The technical report does not provide additional naming rationale beyond this combination.
Summary
VibeThinker-3B represents a meaningful shift in how the AI community thinks about model efficiency and post-training methodology. Here are the key takeaways:
- Training methodology outweighs raw scale. Weibo’s Spectrum-to-Signal pipeline — combining SFT with GRPO — extracted frontier-level performance from an older Qwen2.5-Coder-3B base, proving that careful post-training can rival massive parameter counts for specific tasks.
- The 3B parameter tier is viable for production. With 94.3 on AIME 2026 and 80.2 on LiveCodeBench v6, VibeThinker-3B competes with models over 200 times its size on verifiable reasoning and coding tasks.
- Open source under MIT license removes barriers. Developers can deploy, modify, and commercialize the model freely, making it attractive for cost-sensitive applications and privacy-focused deployments.
- Benchmarks do not equal general capability. The model excels at narrow, verifiable tasks but cannot replace large models for complex, multi-domain work requiring broad knowledge and extended context.
- The base model choice is counterintuitive but instructive. Building on Qwen2.5 rather than Qwen3 demonstrates that newer is not always better when post-training is executed rigorously.
For developers and teams evaluating local AI solutions, VibeThinker-3B deserves serious consideration. Download the weights from Hugging Face, run the benchmarks yourself, and decide whether a 3B specialist fits your pipeline. The results may surprise you.