DEV.co
Open-Source LLM · WeiboAI

VibeThinker-3B

VibeThinker-3B is a 3-billion-parameter open-source LLM from WeiboAI optimized for verifiable reasoning tasks like mathematics, competitive coding, and STEM problems. It uses a post-training pipeline called Spectrum-to-Signal Principle (SSP) combining supervised fine-tuning with reinforcement learning on verifiable rewards. The model achieves competitive performance on reasoning benchmarks (e.g., 76.4% on IMO-AnswerBench, 96.1% on recent LeetCode contests) despite its small size, and is released under the MIT License for unrestricted use. It is not suitable for tool-calling, API orchestration, or open-domain knowledge tasks.

Source: HuggingFace — huggingface.co/WeiboAI/VibeThinker-3B
3.1B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
79.3k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperWeiboAI
Parameters3.1B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads79.3k
Likes779
Last updated2026-06-30
SourceWeiboAI/VibeThinker-3B

What VibeThinker-3B is

VibeThinker-3B is built on Qwen2.5-Coder-3B and trained with a four-stage pipeline: (1) curriculum-based two-stage SFT covering math, code, STEM, and instruction-following; (2) multi-domain reasoning RL using MaxEnt-Guided Policy Optimization (MGPO) applied sequentially to math, code, and STEM with a 64K context window; (3) offline self-distillation filtering high-quality trajectories back into a student model; (4) instruct RL for user-facing controllability. The model uses bfloat16 precision and is distributed via Hugging Face in safetensors format. Inference is recommended via vLLM (0.10.1+) or SGLang (0.4.9.post6+). Context length is unknown. The model card explicitly warns against tool-calling and agent-based programming use.

Quickstart

Run VibeThinker-3B locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="WeiboAI/VibeThinker-3B")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Competitive Math Problem Solving

VibeThinker-3B achieves 76.4% on IMO-AnswerBench and improves to 80.6% with test-time scaling (CLR). It is well-suited for evaluating or training systems on International Mathematical Olympiad-level problems and other math verification tasks where answers can be objectively checked.

Algorithmic Coding Tasks (LeetCode-style)

The model passes 123/128 (96.1%) recent LeetCode weekly/biweekly contest submissions (Python). Ideal for competitive programming, algorithm optimization, and code generation where correctness is verifiable through test cases.

Cost-Effective Reasoning Inference

At 3B parameters, VibeThinker-3B runs on modest hardware (single GPU with 8–16 GB VRAM) while matching or approaching frontier model reasoning performance on verifiable tasks. Suitable for cost-conscious deployments where reasoning capability is the primary requirement.

Running & fine-tuning it

ESTIMATE: 6–12 GB VRAM for inference in bfloat16 on a single consumer GPU (e.g., RTX 4060, A6000). Full fine-tuning requires 24–40 GB VRAM; LoRA/QLoRA fine-tuning reduces to 8–16 GB. Batch inference and longer contexts (max_tokens=60K–100K for evaluation) may require 16–24 GB. Exact requirements depend on batch size, context length, and inference framework (vLLM, SGLang, or transformers). Test on your target hardware.

No explicit LoRA or QLoRA guidance provided in the model card. The model uses bfloat16 and was trained via curriculum SFT + RL; fine-tuning would likely benefit from parameter-efficient methods (LoRA) to preserve verifiable reasoning capabilities without catastrophic forgetting. A learning-potential score was used during self-distillation, suggesting that selective high-quality data may be key. Recommend reviewing the arXiv paper (2606.16140) for training details before attempting significant fine-tuning.

When to avoid it — and what to weigh

  • Tool Calling and Agent Orchestration — The model card explicitly states it was not trained on tool-calling or agent-based programming data and recommends against function calling, API orchestration, or autonomous coding agents. Expect poor performance in these scenarios.
  • Open-Domain Knowledge and General Dialogue — VibeThinker-3B prioritizes verifiable reasoning over broad factual coverage. For open-domain QA, general knowledge retrieval, and diverse conversational tasks, larger general-purpose models are more suitable.
  • Long-Context or Multi-Document Reasoning — Context length is unknown and not provided in model documentation. No evidence of optimization for very long contexts or multi-document retrieval. Requires manual testing or review of technical report.
  • Production Deployments Without Validation — The model is cutting-edge research (June 2026 release) with limited real-world production history. Recommend thorough evaluation on your specific problem domain and edge cases before full deployment.

License & commercial use

VibeThinker-3B is licensed under the MIT License, a permissive OSI-approved license. The model repository is MIT-licensed. No use restrictions are stated in the available data.

MIT License permits commercial use, modification, and distribution without explicit permission, provided the license and copyright notice are retained. No gating, commercial-use restrictions, or special terms are noted in the provided data. However, verify the actual license file in the HuggingFace/GitHub repository and consult legal counsel for large-scale production deployments.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No security audit, adversarial testing, or red-teaming results are described in the provided data. As a reasoning-focused model, standard considerations apply: verify model outputs before deployment, be aware of potential for hallucination or incorrect reasoning on out-of-distribution inputs, and test for prompt injection or jailbreak vectors in production. No claimed protections against model extraction, poisoning, or adversarial examples. Assume standard LLM security hygiene is required.

Alternatives to consider

Qwen2.5-Coder-3B (base model)

VibeThinker-3B is a fine-tuned variant of Qwen2.5-Coder-3B. If you need general-purpose coding without heavy reasoning optimization, the base model is lighter and may be sufficient.

DeepSeek-V3.2 (671B) or Kimi K2.5 (1T)

These are larger frontier models that outperform VibeThinker on some reasoning benchmarks (78.3% and 81.8% on IMO-AnswerBench vs. 76.4%) but require significantly more compute. Consider if hardware budget allows.

VibeThinker-1.5B

Predecessor in the VibeThinker series; smaller and faster but with lower reasoning performance. Suitable for even more resource-constrained deployments.

Software development agency

Ship VibeThinker-3B with senior software developers

Start with a small proof-of-concept on your reasoning use case. Download from Hugging Face, test inference via vLLM or SGLang, and validate outputs before production. Contact our AI engineering team for guidance on fine-tuning, serving, or integration into your system.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

VibeThinker-3B FAQ

Can I use VibeThinker-3B for production commercial applications?
Yes, the MIT License permits commercial use without explicit permission, provided the license and copyright notice are retained. However, verify the actual license in the repository and consult legal counsel, especially for high-stakes applications. The model is research-grade (June 2026 release); validate thoroughly on your use case before production deployment.
What are the minimum hardware requirements for inference?
Approximately 6–12 GB VRAM in bfloat16 on a single consumer GPU (e.g., RTX 4060). For longer contexts or batch inference, 16–24 GB may be needed. Exact requirements depend on your framework (vLLM, SGLang, or Transformers) and batch size. Test on your target hardware before committing.
Why is the model not suitable for API calls or function calling?
The model card explicitly states it was not trained on tool-calling or agent-based programming data. It is optimized for verifiable reasoning (math, code algorithms) where the answer can be checked directly, not for orchestrating external tools or APIs. Use larger general-purpose models for agent tasks.
What is the context length of VibeThinker-3B?
Unknown. The provided data does not specify the context length. The model was trained with a single 64K window during RL and evaluation uses up to 60K–100K max_tokens for hard problems, but the actual supported context length is not stated. Review the technical report (arXiv:2606.16140) or test on your target hardware.

Custom software development services

Adopting VibeThinker-3B is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy VibeThinker-3B?

Start with a small proof-of-concept on your reasoning use case. Download from Hugging Face, test inference via vLLM or SGLang, and validate outputs before production. Contact our AI engineering team for guidance on fine-tuning, serving, or integration into your system.