# VibeThinker 3B: Beats Claude Opus & OpenAI at 3B Parameters

WeiboAI's VibeThinker-3B punches way above its weight class — matching or exceeding models 200-300x larger on reasoning benchmarks. MIT licensed, open source, but with one major caveat.

## 🚨 The Big Caveat First: No Tool Calling

Before diving into the impressive benchmarks, here's the critical limitation that every user needs to know: **VibeThinker-3B was never trained on tool-calling or agent-based programming data.** The authors explicitly state on the model card that it is *not recommended for function calling, API orchestration, or autonomous coding agents*.

**⚠️ Not Built for Agents:** If you're looking for a small model that can call tools, use APIs, or act as an autonomous agent — this is not it. The creators recommend using it specifically for competitive-style math, coding (LeetCode), STEM reasoning, and any task where the answer can be verified.

That said, for its *intended* use case — verifiable reasoning tasks with clear feedback signals — VibeThinker-3B is nothing short of remarkable.

## 📊 Benchmark Performance That Shouldn't Be Possible

At just **3 billion parameters** (a dense Qwen2 architecture, not MoE), VibeThinker-3B achieves scores that put it in the performance band of first-tier reasoning systems — models that are *hundreds of times larger*.

94.3

AIME26 Score

97.1

AIME26 + CLR

80.2

LiveCodeBench v6

96.1%

LeetCode Acceptance

76.4

IMO AnswerBench

93.4

IFEval

Let's break down what these numbers actually mean in context.

### AIME26: Near-Perfect Math Reasoning

VibeThinker-3B scores **94.3 on AIME26** — a competition-level mathematics benchmark. With **Claim-Level Reliability Assessment (CLR)**, a novel test-time scaling strategy, it reaches **97.1**. These are frontier-level scores that put it in the same performance band as models like DeepSeek V3.2, GLM-5, and Gemini 3 Pro.

### LiveCodeBench v6: Competitive Coding at Scale

On LiveCodeBench v6 — a coding benchmark testing real-world algorithmic problem solving — VibeThinker-3B achieves **80.2 Pass@1**. For out-of-distribution generalization on recent unseen LeetCode contests (April–May 2026), it passes **123 out of 128 first-attempt submissions** — a **96.1% acceptance rate**.

### IMO AnswerBench: Olympiad-Level Math

The International Mathematical Olympiad benchmark (400 problems) is one of the hardest math reasoning tests available. VibeThinker-3B scores **76.4**, improving to **80.6 with CLR** — rivaling models 200-300x its size.

## 📈 How It Compares to Flagship Models

The most striking aspect of VibeThinker-3B is how its performance scales relative to model size. Here's how it compares to much larger systems on IMO AnswerBench:

| **Model** | **Params** | **IMO AnswerBench** | **IMO + CLR**|
--- | --- | --- | ---
| **VibeThinker-3B** | **3B** | **76.4** | **80.6**|
| DeepSeek V3.2 | 671B | 78.3 | —|
| GLM-5 | 744B | 82.5 | —|
| Kimi K2.5 | 1T | 81.8 | —|

The key insight: VibeThinker-3B achieves **nearly the same reasoning score as models that are 250-300x larger**. With CLR boosting it to 80.6, it's within striking distance of GLM-5's 82.5 — on a model that fits in roughly **6 GB of VRAM** (BF16) or even less when quantized to GGUF.

## 🧠 The Training Methodology: Spectrum-to-Signal Principle

VibeThinker-3B builds on the **Spectrum-to-Signal Principle (SSP)**, a post-training paradigm introduced in its smaller sibling VibeThinker-1.5B. The approach is systematic and methodical:

### Stage 1: Curriculum-based Two-Stage SFT

Stage 1 builds broad capability coverage across math, code, STEM reasoning, general dialogue, and instruction following. Stage 2 shifts focus toward harder and longer-horizon reasoning samples. A technique called **Diversity-Exploring Distillation** preserves multiple valid solution paths.

### Stage 2: Multi-Domain Reasoning RL

Uses **MaxEnt-Guided Policy Optimization (MGPO)**. Reinforcement learning is applied sequentially to math, code, and STEM reasoning tasks. Training uses a single 64K long-context window to preserve complete long-horizon reasoning trajectories.

### Stage 3: Offline Self-Distillation

High-quality trajectories from Math, Code, and STEM RL checkpoints are filtered and distilled back into a unified student model. A **learning-potential score** prioritizes traces that are correct but not yet well modeled by the student.

### Stage 4: Instruct RL

Final stage improves controllability on user-facing prompts. Rule-based validators and rubric-based reward models are used for format-sensitive and open-ended instruction data.

## 🔬 Key Research Insight: The Parametric Compression-Coverage Hypothesis

Beyond the benchmarks, VibeThinker-3B introduces an interesting theoretical framework. The authors propose the **Parametric Compression-Coverage Hypothesis**:

"Verifiable reasoning is closer to a highly compressible, parameter-dense capability, centered on multi-step reasoning, constraint satisfaction, self-correction, and answer verification. When the task space is sufficiently structured and feedback signals are sufficiently reliable, compact models may also carry near-frontier reasoning capabilities."
— VibeThinker Technical Report (arXiv:2606.16140)

In simpler terms: **reasoning is compressible**. You don't need massive parameter counts to be great at math, code, and STEM problems — you need the right training methodology. Open-domain knowledge and general-purpose dialogue, on the other hand, *do* benefit from scale.

This means compact models shouldn't be viewed merely as "cheap compromises" — they can be a **complementary path** toward frontier-level performance in specific capability domains.

## 📋 Technical Specifications

| **Property** | **Value**|
--- | ---
| Parameter Count | 3B (dense)|
| Architecture | Qwen2|
| Base Model | Qwen/Qwen2.5-Coder-3B|
| Hidden Size | 2,048|
| Layers | 36|
| Attention Heads | 16 (16 query, 2 KV)|
| Vocabulary | 151,936|
| Context Window | 131,072 tokens|
| RoPE Theta | 1,000,000|
| Sliding Window | 32,768 (disabled by default)|
| Activation | SiLU|
| Embeddings | Tied|
| License | MIT|
| Downloads | 32,385+|
| Likes | 613|
| Quantization | GGUF, Safetensors|
| Release Date | June 12, 2026|

## ⚙️ Quick Start

For best inference performance, use **vLLM 0.10.1+** or **SGLang 0.4.9+:**

```
# Install vLLM
pip install vllm

# Start the server
vllm serve "WeiboAI/VibeThinker-3B"

# Call via OpenAI-compatible API
curl -X POST "http://localhost:8000/v1/chat/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "WeiboAI/VibeThinker-3B",
"messages": [{"role": "user", "content": "Solve: 25 * 37 + 13"}],
"temperature": 1.0,
"top_p": 0.95
}'
```

**Recommended inference settings:** temperature=1.0, top_p=0.95, top_k=-1, max_tokens up to 60K–100K for IMO-level problems.

## 🎯 When to Use (and When Not To)

### ✅ Great For

- Competitive programming (LeetCode-style problems)

- Mathematical reasoning and proof generation

- STEM problem solving with clear answers

- Instruction-following and format-sensitive tasks (IFEval: 93.4)

- Edge/on-device deployment where 6GB VRAM is a hard limit

- Benchmark evaluation and research on small model reasoning

### ❌ Not Suitable For

- Tool calling or function calling

- API orchestration or agent workflows

- Autonomous coding agents

- Broad open-domain knowledge QA

- General-purpose conversational AI

- Tasks requiring extensive world knowledge

## 🏁 Conclusion

VibeThinker-3B is a bold experiment in parameter efficiency that delivers real results. At **3 billion parameters** with an **MIT license**, it matches the reasoning performance of models 200–300x its size on verifiable tasks like math, coding, and STEM — and its **Claim-Level Reliability Assessment** technique pushes it even closer to true frontier performance.

The model makes a compelling case that for capability domains with clear feedback and verification mechanisms, **small language models are not just a deployment compromise — they're a genuinely viable research trajectory toward frontier-level reasoning**.

Just remember: no tool calling. For that, you still need the big guns. But for pure reasoning on structured tasks? VibeThinker-3B might just be the most efficient model on the planet right now.

"Small models should not be viewed merely as a compromise for reducing deployment costs. For capability domains with clear feedback and verification mechanisms, SLMs emerge as a promising research trajectory toward frontier-level performance."
— WeiboAI VibeThinker Team

[🤗 HuggingFace](https://huggingface.co/WeiboAI/VibeThinker-3B)
[📄 arXiv Paper](https://arxiv.org/abs/2606.16140)
[🐙 GitHub](https://github.com/WeiboAI/VibeThinker)