OMCHOKSI108/VibeThinker-3B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 23, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

VibeThinker-3B by WeiboAI is a 3-billion-parameter dense reasoning model built upon Qwen2.5-Coder-3B, post-trained with an upgraded Spectrum-to-Signal (SSP) pipeline. It is specifically designed for tasks requiring reliable verification signals, such as mathematical reasoning, competitive programming, and STEM reasoning. This model achieves frontier-level performance on challenging reasoning benchmarks, demonstrating high-density reasoning ability in a compact 3B parameter size with a 32K context length.

Loading preview...

VibeThinker-3B: Ultra-Efficient Reasoning Model

VibeThinker-3B, developed by WeiboAI, is a 3-billion-parameter dense reasoning model based on Qwen2.5-Coder-3B. It leverages an upgraded Spectrum-to-Signal (SSP) pipeline to achieve exceptional performance on tasks with verifiable solutions, despite its compact size.

Key Capabilities & Performance

  • Ultra-Efficient Frontier-Level Reasoning: Approaches the performance of much larger reasoning systems on challenging benchmarks with only 3B parameters.
  • Outstanding Benchmark Results: Achieves 94.3 on AIME26, 89.3 on HMMT25, 80.2 Pass@1 on LiveCodeBench v6, and a 96.1% acceptance rate on recent unseen LeetCode contests.
  • Inference-Time Scaling with CLR: Introduces Claim-Level Reliability Assessment (CLR) to further boost performance on math benchmarks, raising AIME26 to 97.1 and HMMT25 to 95.4.
  • Robust Out-of-Distribution Performance: Demonstrated strong performance on unseen LeetCode contests, passing 123 out of 128 first-attempt submissions.

Training Pipeline Highlights

The model's training follows the Spectrum-to-Signal Principle (SSP), involving a curriculum-based two-stage Supervised Fine-Tuning (SFT), Multi-domain Reasoning Reinforcement Learning (RL) with a 64K context window, Offline Self-Distillation, and Instruct RL for improved controllability.

Good For

  • Mathematical reasoning (e.g., AIME, HMMT)
  • Competitive programming (e.g., LeetCode, LiveCodeBench)
  • STEM reasoning tasks
  • Instruction-following with explicit constraints

Limitations

  • Not recommended for tool-calling, API orchestration, or autonomous coding agents.
  • Larger general-purpose models may be more suitable for open-domain knowledge tasks.

Finetunes
1 models